The new Power7 processor has 4 way SMT. This 4 way SMT benefits from
dynamic power updates that arch_scale_smt_power was designed to provide.
The first patch fixes a generic scheduler bug necessary for arch_scale_smt
to properly function. The second patch implements arch_scale_smt_power
for powerpc, and in particular for Power7 processors.
---
Gautham R Shenoy (1):
sched: Fix the place where group powers are updated.
Joel Schopp (1):
powerpc: implement arch_scale_smt_power for Power7
arch/powerpc/kernel/smp.c | 41 +++++++++++++++++++++++++++++++++++++++++
kernel/sched.c | 7 +++----
kernel/sched_features.h | 2 +-
3 files changed, 45 insertions(+), 5 deletions(-)
From: Gautham R Shenoy <[email protected]>
We want to update the sched_group_powers when balance_cpu == this_cpu.
Currently the group powers are updated only if the balance_cpu is the first
CPU in the local group. But balance_cpu = this_cpu could also be the first
idle cpu in the group. Hence fix the place where the group powers are updated.
Signed-off-by: Gautham R Shenoy <[email protected]>
Signed-off-by: Joel Schopp <[email protected]>
---
kernel/sched.c | 7 +++----
1 files changed, 3 insertions(+), 4 deletions(-)
diff --git a/kernel/sched.c b/kernel/sched.c
index 281da29..5d2a451 100644
--- a/kernel/sched.c
+++ b/kernel/sched.c
@@ -3721,11 +3721,8 @@ static inline void update_sg_lb_stats(struct sched_domain *sd,
unsigned long sum_avg_load_per_task;
unsigned long avg_load_per_task;
- if (local_group) {
+ if (local_group)
balance_cpu = group_first_cpu(group);
- if (balance_cpu == this_cpu)
- update_group_power(sd, this_cpu);
- }
/* Tally up the load of all CPUs in the group */
sum_avg_load_per_task = avg_load_per_task = 0;
@@ -3773,6 +3770,8 @@ static inline void update_sg_lb_stats(struct sched_domain *sd,
return;
}
+ update_group_power(sd, this_cpu);
+
/* Adjust by relative CPU power of the group */
sgs->avg_load = (sgs->group_load * SCHED_LOAD_SCALE) / group->cpu_power;
On Power7 processors running in SMT4 mode with 2, 3, or 4 idle threads
there is performance benefit to idling the higher numbered threads in
the core.
This patch implements arch_scale_smt_power to dynamically update smt
thread power in these idle cases in order to prefer threads 0,1 over
threads 2,3 within a core.
Signed-off-by: Joel Schopp <[email protected]>
---
Index: linux-2.6.git/arch/powerpc/kernel/smp.c
===================================================================
--- linux-2.6.git.orig/arch/powerpc/kernel/smp.c
+++ linux-2.6.git/arch/powerpc/kernel/smp.c
@@ -617,3 +617,44 @@ void __cpu_die(unsigned int cpu)
smp_ops->cpu_die(cpu);
}
#endif
+
+static inline int thread_in_smt4core(int x)
+{
+ return x % 4;
+}
+unsigned long arch_scale_smt_power(struct sched_domain *sd, int cpu)
+{
+ int cpu2;
+ int idle_count = 0;
+
+ struct cpumask *cpu_map = sched_domain_span(sd);
+
+ unsigned long weight = cpumask_weight(cpu_map);
+ unsigned long smt_gain = sd->smt_gain;
+
+ if (cpu_has_feature(CPU_FTRS_POWER7) && weight == 4) {
+ for_each_cpu(cpu2, cpu_map) {
+ if (idle_cpu(cpu2))
+ idle_count++;
+ }
+
+ /* the following section attempts to tweak cpu power based
+ * on current idleness of the threads dynamically at runtime
+ */
+ if (idle_count == 2 || idle_count == 3 || idle_count == 4) {
+ if (thread_in_smt4core(cpu) == 0 ||
+ thread_in_smt4core(cpu) == 1) {
+ /* add 75 % to thread power */
+ smt_gain += (smt_gain >> 1) + (smt_gain >> 2);
+ } else {
+ /* subtract 75 % to thread power */
+ smt_gain = smt_gain >> 2;
+ }
+ }
+ }
+ /* default smt gain is 1178, weight is # of SMT threads */
+ smt_gain /= weight;
+
+ return smt_gain;
+
+}
Index: linux-2.6.git/kernel/sched_features.h
===================================================================
--- linux-2.6.git.orig/kernel/sched_features.h
+++ linux-2.6.git/kernel/sched_features.h
@@ -107,7 +107,7 @@ SCHED_FEAT(CACHE_HOT_BUDDY, 1)
/*
* Use arch dependent cpu power functions
*/
-SCHED_FEAT(ARCH_POWER, 0)
+SCHED_FEAT(ARCH_POWER, 1)
SCHED_FEAT(HRTICK, 0)
SCHED_FEAT(DOUBLE_TICK, 0)
On Wed, 2010-01-20 at 14:04 -0600, Joel Schopp wrote:
> On Power7 processors running in SMT4 mode with 2, 3, or 4 idle threads
> there is performance benefit to idling the higher numbered threads in
> the core.
So this is an actual performance improvement, not only power savings?
> This patch implements arch_scale_smt_power to dynamically update smt
> thread power in these idle cases in order to prefer threads 0,1 over
> threads 2,3 within a core.
>
> Signed-off-by: Joel Schopp <[email protected]>
> ---
> Index: linux-2.6.git/arch/powerpc/kernel/smp.c
> ===================================================================
> --- linux-2.6.git.orig/arch/powerpc/kernel/smp.c
> +++ linux-2.6.git/arch/powerpc/kernel/smp.c
> @@ -617,3 +617,44 @@ void __cpu_die(unsigned int cpu)
> smp_ops->cpu_die(cpu);
> }
> #endif
> +
> +static inline int thread_in_smt4core(int x)
> +{
> + return x % 4;
> +}
> +unsigned long arch_scale_smt_power(struct sched_domain *sd, int cpu)
> +{
> + int cpu2;
> + int idle_count = 0;
> +
> + struct cpumask *cpu_map = sched_domain_span(sd);
> +
> + unsigned long weight = cpumask_weight(cpu_map);
> + unsigned long smt_gain = sd->smt_gain;
> +
> + if (cpu_has_feature(CPU_FTRS_POWER7) && weight == 4) {
> + for_each_cpu(cpu2, cpu_map) {
> + if (idle_cpu(cpu2))
> + idle_count++;
> + }
> +
> + /* the following section attempts to tweak cpu power based
> + * on current idleness of the threads dynamically at runtime
> + */
> + if (idle_count == 2 || idle_count == 3 || idle_count == 4) {
> + if (thread_in_smt4core(cpu) == 0 ||
> + thread_in_smt4core(cpu) == 1) {
> + /* add 75 % to thread power */
> + smt_gain += (smt_gain >> 1) + (smt_gain >> 2);
> + } else {
> + /* subtract 75 % to thread power */
> + smt_gain = smt_gain >> 2;
> + }
> + }
> + }
> + /* default smt gain is 1178, weight is # of SMT threads */
> + smt_gain /= weight;
> +
> + return smt_gain;
> +
> +}
This looks to suffer significant whitespace damage.
The design goal for smt_power was to be able to actually measure the
processing gains from smt and feed that into the scheduler, not really
placement tricks like this.
Now I also heard AMD might want to have something similar to this,
something to do with powerlines and die layout.
I'm not sure playing games with cpu_power is the best or if simply
moving tasks to lower numbered cpus using an SD_flag is the best
solution for these kinds of things.
> Index: linux-2.6.git/kernel/sched_features.h
> ===================================================================
> --- linux-2.6.git.orig/kernel/sched_features.h
> +++ linux-2.6.git/kernel/sched_features.h
> @@ -107,7 +107,7 @@ SCHED_FEAT(CACHE_HOT_BUDDY, 1)
> /*
> * Use arch dependent cpu power functions
> */
> -SCHED_FEAT(ARCH_POWER, 0)
> +SCHED_FEAT(ARCH_POWER, 1)
>
> SCHED_FEAT(HRTICK, 0)
> SCHED_FEAT(DOUBLE_TICK, 0)
And you just wrecked x86 ;-)
It has an smt_power implementation that tries to measure smt gains using
aperf/mperf, trouble is that this represents the actual performance not
the capacity. This has the problem that when idle it represents 0
capacity and will not attract work.
Coming up with something that actually works there is on the todo list,
I was thinking perhaps temporal maximums from !idle.
So if you want to go with this, you'll need to stub out
arch/x86/kernel/cpu/sched.c
> On Power7 processors running in SMT4 mode with 2, 3, or 4 idle threads
> there is performance benefit to idling the higher numbered threads in
> the core.
>
> This patch implements arch_scale_smt_power to dynamically update smt
> thread power in these idle cases in order to prefer threads 0,1 over
> threads 2,3 within a core.
>
> Signed-off-by: Joel Schopp <[email protected]>
> ---
> Index: linux-2.6.git/arch/powerpc/kernel/smp.c
> ===================================================================
> --- linux-2.6.git.orig/arch/powerpc/kernel/smp.c
> +++ linux-2.6.git/arch/powerpc/kernel/smp.c
> @@ -617,3 +617,44 @@ void __cpu_die(unsigned int cpu)
> smp_ops->cpu_die(cpu);
> }
> #endif
> +
> +static inline int thread_in_smt4core(int x)
> +{
> + return x % 4;
> +}
> +unsigned long arch_scale_smt_power(struct sched_domain *sd, int cpu)
> +{
> + int cpu2;
> + int idle_count = 0;
> +
> + struct cpumask *cpu_map = sched_domain_span(sd);
> +
> + unsigned long weight = cpumask_weight(cpu_map);
> + unsigned long smt_gain = sd->smt_gain;
> +
> + if (cpu_has_feature(CPU_FTRS_POWER7) && weight == 4) {
I think we should avoid using cpu_has_feature like this. It's better to
create a new feature and add it to POWER7 in the cputable, then check
for that here.
The way that it is now, I think any CPU that has superset of the POWER7
features, will be true here. This is not what we want.
> + for_each_cpu(cpu2, cpu_map) {
> + if (idle_cpu(cpu2))
> + idle_count++;
> + }
> +
> + /* the following section attempts to tweak cpu power based
> + * on current idleness of the threads dynamically at runtime
> + */
> + if (idle_count == 2 || idle_count == 3 || idle_count == 4) {
> + if (thread_in_smt4core(cpu) == 0 ||
> + thread_in_smt4core(cpu) == 1) {
> + /* add 75 % to thread power */
> + smt_gain += (smt_gain >> 1) + (smt_gain >> 2);
> + } else {
> + /* subtract 75 % to thread power */
> + smt_gain = smt_gain >> 2;
> + }
> + }
> + }
> + /* default smt gain is 1178, weight is # of SMT threads */
> + smt_gain /= weight;
This results in a PPC div, when most of the time it's going to be a
power of two divide. You've optimised the divides a few lines above
this, but not this one. Some consistency would be good.
Mikey
> +
> + return smt_gain;
> +
> +}
> Index: linux-2.6.git/kernel/sched_features.h
> ===================================================================
> --- linux-2.6.git.orig/kernel/sched_features.h
> +++ linux-2.6.git/kernel/sched_features.h
> @@ -107,7 +107,7 @@ SCHED_FEAT(CACHE_HOT_BUDDY, 1)
> /*
> * Use arch dependent cpu power functions
> */
> -SCHED_FEAT(ARCH_POWER, 0)
> +SCHED_FEAT(ARCH_POWER, 1)
>
> SCHED_FEAT(HRTICK, 0)
> SCHED_FEAT(DOUBLE_TICK, 0)
>
>
> _______________________________________________
> Linuxppc-dev mailing list
> [email protected]
> https://lists.ozlabs.org/listinfo/linuxppc-dev
>
On Wed, 2010-01-20 at 14:04 -0600, Joel Schopp wrote:
> On Power7 processors running in SMT4 mode with 2, 3, or 4 idle threads
> there is performance benefit to idling the higher numbered threads in
> the core.
>
> This patch implements arch_scale_smt_power to dynamically update smt
> thread power in these idle cases in order to prefer threads 0,1 over
> threads 2,3 within a core.
>
> Signed-off-by: Joel Schopp <[email protected]>
So I'll leave Peter deal with the scheduler aspects and will focus on
details :-)
> ---
> Index: linux-2.6.git/arch/powerpc/kernel/smp.c
> ===================================================================
> --- linux-2.6.git.orig/arch/powerpc/kernel/smp.c
> +++ linux-2.6.git/arch/powerpc/kernel/smp.c
> @@ -617,3 +617,44 @@ void __cpu_die(unsigned int cpu)
> smp_ops->cpu_die(cpu);
> }
> #endif
> +
> +static inline int thread_in_smt4core(int x)
> +{
> + return x % 4;
> +}
Needs a whitespace here though I don't really like the above. Any reason
why you can't use the existing cpu_thread_in_core() ?
> +unsigned long arch_scale_smt_power(struct sched_domain *sd, int cpu)
> +{
> + int cpu2;
> + int idle_count = 0;
> +
> + struct cpumask *cpu_map = sched_domain_span(sd);
> +
> + unsigned long weight = cpumask_weight(cpu_map);
> + unsigned long smt_gain = sd->smt_gain;
More whitespace damage above.
> + if (cpu_has_feature(CPU_FTRS_POWER7) && weight == 4) {
> + for_each_cpu(cpu2, cpu_map) {
> + if (idle_cpu(cpu2))
> + idle_count++;
> + }
I'm not 100% sure about the use of the CPU feature above. First I wonder
if the right approach is to instead do something like
if (!cpu_has_feature(...) !! weigth < 4)
return default_scale_smt_power(sd, cpu);
Though we may be better off using a ppc_md. hook here to avoid
calculating the weight etc... on processors that don't need any
of that.
I also dislike your naming. I would suggest you change cpu_map to
sibling_map() and cpu2 to sibling (or just c). One thing I wonder is how
sure we are that sched_domain_span() is always going to give us the
threads btw ? If we introduce another sched domain level for NUMA
purposes can't we get confused ?
Also, how hot is this code path ?
> + /* the following section attempts to tweak cpu power based
> + * on current idleness of the threads dynamically at runtime
> + */
> + if (idle_count == 2 || idle_count == 3 || idle_count == 4) {
if (idle_count > 1) ? :-)
> + if (thread_in_smt4core(cpu) == 0 ||
> + thread_in_smt4core(cpu) == 1) {
int thread = cpu_thread_in_core(cpu);
if (thread < 2)
...
> + /* add 75 % to thread power */
> + smt_gain += (smt_gain >> 1) + (smt_gain >> 2);
> + } else {
> + /* subtract 75 % to thread power */
> + smt_gain = smt_gain >> 2;
> + }
> + }
> + }
> + /* default smt gain is 1178, weight is # of SMT threads */
> + smt_gain /= weight;
> +
> + return smt_gain;
Cheers,
Ben.
> +}
> Index: linux-2.6.git/kernel/sched_features.h
> ===================================================================
> --- linux-2.6.git.orig/kernel/sched_features.h
> +++ linux-2.6.git/kernel/sched_features.h
> @@ -107,7 +107,7 @@ SCHED_FEAT(CACHE_HOT_BUDDY, 1)
> /*
> * Use arch dependent cpu power functions
> */
> -SCHED_FEAT(ARCH_POWER, 0)
> +SCHED_FEAT(ARCH_POWER, 1)
>
> SCHED_FEAT(HRTICK, 0)
> SCHED_FEAT(DOUBLE_TICK, 0)
>
> > On Power7 processors running in SMT4 mode with 2, 3, or 4 idle threads
> > there is performance benefit to idling the higher numbered threads in
> > the core.
>
> So this is an actual performance improvement, not only power savings?
It's primarily a performance improvement. Any power/energy savings
would be a bonus.
Mikey
>> + if (cpu_has_feature(CPU_FTRS_POWER7) && weight == 4) {
>>
>
> I think we should avoid using cpu_has_feature like this. It's better to
> create a new feature and add it to POWER7 in the cputable, then check
> for that here.
>
> The way that it is now, I think any CPU that has superset of the POWER7
> features, will be true here. This is not what we want.
>
Any ideas for what to call this feature? ASYM_SMT4 ?
>
>> + smt_gain /= weight;
>>
>
> This results in a PPC div, when most of the time it's going to be a
> power of two divide. You've optimised the divides a few lines above
> this, but not this one. Some consistency would be good.
>
>
I can turn that into a conditional branch (case statement) with a shift
for the common 1,2,4 cases which should cover all procs available today
falling back to a divide for any theoretical future processors that do
other numbers of threads.
>> +
>> +static inline int thread_in_smt4core(int x)
>> +{
>> + return x % 4;
>> +}
>>
>
> Needs a whitespace here though I don't really like the above. Any reason
> why you can't use the existing cpu_thread_in_core() ?
>
I will change it to cpu_thread_in_core()
>
>> +unsigned long arch_scale_smt_power(struct sched_domain *sd, int cpu)
>> +{
>> + int cpu2;
>> + int idle_count = 0;
>> +
>> + struct cpumask *cpu_map = sched_domain_span(sd);
>> +
>> + unsigned long weight = cpumask_weight(cpu_map);
>> + unsigned long smt_gain = sd->smt_gain;
>>
>
> More whitespace damage above.
>
You are better than checkpatch.pl, will fix.
>
>> + if (cpu_has_feature(CPU_FTRS_POWER7) && weight == 4) {
>> + for_each_cpu(cpu2, cpu_map) {
>> + if (idle_cpu(cpu2))
>> + idle_count++;
>> + }
>>
>
> I'm not 100% sure about the use of the CPU feature above. First I wonder
> if the right approach is to instead do something like
>
> if (!cpu_has_feature(...) !! weigth < 4)
> return default_scale_smt_power(sd, cpu);
>
> Though we may be better off using a ppc_md. hook here to avoid
> calculating the weight etc... on processors that don't need any
> of that.
>
> I also dislike your naming. I would suggest you change cpu_map to
> sibling_map() and cpu2 to sibling (or just c). One thing I wonder is how
> sure we are that sched_domain_span() is always going to give us the
> threads btw ? If we introduce another sched domain level for NUMA
> purposes can't we get confused ?
>
Right now it's 100% always giving us threads. My development version of
the patch had a BUG_ON() to check this. I expect this to stay the case
in the future as the name of the function is arch_scale_smt_power(),
which clearly denotes threads are expected.
I am not stuck on the names, I'll change it to sibling instead of cpu2
and sibling_map instead of cpu_map. It seems clear to me either way.
As for testing the ! case it seems funcationally equivalent, and mine
seems less confusing.
Having a ppc.md hook with exactly 1 user is pointless, especially since
you'll still have to calculate the weight with the ability to
dynamically disable smt.
> Also, how hot is this code path ?
>
It's every load balance, which is to say not hot, but fairly frequent.
I haven't been able to measure an impact from doing very hairy
calculations (without actually changing the weights) here vs not having
it at all in actual end workloads.
>
>> + /* the following section attempts to tweak cpu power based
>> + * on current idleness of the threads dynamically at runtime
>> + */
>> + if (idle_count == 2 || idle_count == 3 || idle_count == 4) {
>>
>
> if (idle_count > 1) ? :-)
>
Yes :) Originally I had done different weightings for each of the 3
cases, which gained in some workloads but regressed some others. But
since I'm not doing that anymore I'll fold it down to > 1
Peter Zijlstra wrote:
> On Wed, 2010-01-20 at 14:04 -0600, Joel Schopp wrote:
>
>> On Power7 processors running in SMT4 mode with 2, 3, or 4 idle threads
>> there is performance benefit to idling the higher numbered threads in
>> the core.
>>
>
> So this is an actual performance improvement, not only power savings?
>
Yes.
>
>
> And you just wrecked x86 ;-)
>
> It has an smt_power implementation that tries to measure smt gains using
> aperf/mperf, trouble is that this represents the actual performance not
> the capacity. This has the problem that when idle it represents 0
> capacity and will not attract work.
>
> Coming up with something that actually works there is on the todo list,
> I was thinking perhaps temporal maximums from !idle.
>
> So if you want to go with this, you'll need to stub out
> arch/x86/kernel/cpu/sched.c
>
OK. Guess I now will have a 3 patch series, with a patch to stub out
the x86 broken version.
Care to take Gautham's bugfix patch (patch 1/2) now, since it just fixes
a bug? You'll need it if you ever try to make the x86 broken version work.
On Wed, 2010-01-20 at 16:44 -0600, Joel Schopp wrote:
>
> Care to take Gautham's bugfix patch (patch 1/2) now, since it just fixes
> a bug? You'll need it if you ever try to make the x86 broken version work.
Sure, I'll take that, thanks!
Commit-ID: 871e35bc9733f273eaf5ceb69bbd0423b58e5285
Gitweb: http://git.kernel.org/tip/871e35bc9733f273eaf5ceb69bbd0423b58e5285
Author: Gautham R Shenoy <[email protected]>
AuthorDate: Wed, 20 Jan 2010 14:02:44 -0600
Committer: Ingo Molnar <[email protected]>
CommitDate: Thu, 21 Jan 2010 13:40:17 +0100
sched: Fix the place where group powers are updated
We want to update the sched_group_powers when balance_cpu == this_cpu.
Currently the group powers are updated only if the balance_cpu is the
first CPU in the local group. But balance_cpu = this_cpu could also be
the first idle cpu in the group. Hence fix the place where the group
powers are updated.
Signed-off-by: Gautham R Shenoy <[email protected]>
Signed-off-by: Joel Schopp <[email protected]>
Signed-off-by: Peter Zijlstra <[email protected]>
LKML-Reference: <1264017764.5717.127.camel@jschopp-laptop>
Signed-off-by: Ingo Molnar <[email protected]>
---
kernel/sched_fair.c | 7 +++----
1 files changed, 3 insertions(+), 4 deletions(-)
diff --git a/kernel/sched_fair.c b/kernel/sched_fair.c
index 0b482f5..22231cc 100644
--- a/kernel/sched_fair.c
+++ b/kernel/sched_fair.c
@@ -2418,11 +2418,8 @@ static inline void update_sg_lb_stats(struct sched_domain *sd,
unsigned long sum_avg_load_per_task;
unsigned long avg_load_per_task;
- if (local_group) {
+ if (local_group)
balance_cpu = group_first_cpu(group);
- if (balance_cpu == this_cpu)
- update_group_power(sd, this_cpu);
- }
/* Tally up the load of all CPUs in the group */
sum_avg_load_per_task = avg_load_per_task = 0;
@@ -2470,6 +2467,8 @@ static inline void update_sg_lb_stats(struct sched_domain *sd,
return;
}
+ update_group_power(sd, this_cpu);
+
/* Adjust by relative CPU power of the group */
sgs->avg_load = (sgs->group_load * SCHED_LOAD_SCALE) / group->cpu_power;
On Wed, 2010-01-20 at 16:09 -0600, Joel Schopp wrote:
> I can turn that into a conditional branch (case statement) with a shift
> for the common 1,2,4 cases which should cover all procs available today
> falling back to a divide for any theoretical future processors that do
> other numbers of threads.
Look at the cputhreads.h implementation ... Today we only support
power-of-two numbers of threads.
Cheers,
Ben.
Benjamin Herrenschmidt wrote:
> On Wed, 2010-01-20 at 16:09 -0600, Joel Schopp wrote:
>
>> I can turn that into a conditional branch (case statement) with a shift
>> for the common 1,2,4 cases which should cover all procs available today
>> falling back to a divide for any theoretical future processors that do
>> other numbers of threads.
>>
>
> Look at the cputhreads.h implementation ... Today we only support
> power-of-two numbers of threads.
>
I've run 3 threads using cpu hotplug to offline 1 of the 4. It's
certainly a stupid idea, but there you go.
On Mon, 2010-01-25 at 11:50 -0600, Joel Schopp wrote:
> > Look at the cputhreads.h implementation ... Today we only support
> > power-of-two numbers of threads.
> >
> I've run 3 threads using cpu hotplug to offline 1 of the 4. It's
> certainly a stupid idea, but there you go.
Oh, you mean you need to use the actual online count ? In which case,
yes, you do indeed need to be a bit more careful... In this case though,
you're probably better off special casing "3" :-)
Cheers,
Ben.
The new Power7 processor has 4 way SMT. This 4 way SMT benefits from
dynamic power updates that arch_scale_smt_power was designed to provide.
The first patch fixes a generic scheduler bug necessary for arch_scale_smt
to properly function. The second patch implements arch_scale_smt_power
for powerpc, and in particular for Power7 processors.
Version 2 changes:
- Drop Patch 1 from the original series since it's in the -tip tree now
- Move enabling the cpu feature into it's own patch (now patch 1)
- Add stubbing out broken x86 implementation to patch 1
- clean up coding style in patch 2
---
Joel Schopp (2):
powerpc: implement arch_scale_smt_power for Power7
sched: enable ARCH_POWER
arch/powerpc/include/asm/cputable.h | 3 +-
arch/powerpc/kernel/smp.c | 52 ++++++++++++++++++++++++++++++++++++
arch/x86/kernel/cpu/sched.c | 6 +---
kernel/sched_features.h | 2 -
4 files changed, 57 insertions(+), 6 deletions(-)
Enable the scheduler feature that allows use of arch_scale_smt_power. Stub out
the broken x86 implementation.
Signed-off-by: Joel Schopp <[email protected]>
---
Index: linux-2.6.git/kernel/sched_features.h
===================================================================
--- linux-2.6.git.orig/kernel/sched_features.h
+++ linux-2.6.git/kernel/sched_features.h
@@ -102,7 +102,7 @@ SCHED_FEAT(CACHE_HOT_BUDDY, 1)
/*
* Use arch dependent cpu power functions
*/
-SCHED_FEAT(ARCH_POWER, 0)
+SCHED_FEAT(ARCH_POWER, 1)
SCHED_FEAT(HRTICK, 0)
SCHED_FEAT(DOUBLE_TICK, 0)
Index: linux-2.6.git/arch/x86/kernel/cpu/sched.c
===================================================================
--- linux-2.6.git.orig/arch/x86/kernel/cpu/sched.c
+++ linux-2.6.git/arch/x86/kernel/cpu/sched.c
@@ -44,11 +44,9 @@ unsigned long arch_scale_freq_power(stru
unsigned long arch_scale_smt_power(struct sched_domain *sd, int cpu)
{
/*
- * aperf/mperf already includes the smt gain
+ * aperf/mperf already includes the smt gain, but represents capacity
+ * as 0 when idle. So for now just return default.
*/
- if (boot_cpu_has(X86_FEATURE_APERFMPERF))
- return SCHED_LOAD_SCALE;
-
return default_scale_smt_power(sd, cpu);
}
On Power7 processors running in SMT4 mode with 2, 3, or 4 idle threads
there is performance benefit to idling the higher numbered threads in
the core.
This patch implements arch_scale_smt_power to dynamically update smt
thread power in these idle cases in order to prefer threads 0,1 over
threads 2,3 within a core.
v2 - Same functionality as v1, better coding style.
Signed-off-by: Joel Schopp <[email protected]>
---
Version 2 addresses style and optimization, same basic functionality
Index: linux-2.6.git/arch/powerpc/kernel/smp.c
===================================================================
--- linux-2.6.git.orig/arch/powerpc/kernel/smp.c
+++ linux-2.6.git/arch/powerpc/kernel/smp.c
@@ -620,3 +620,55 @@ void __cpu_die(unsigned int cpu)
smp_ops->cpu_die(cpu);
}
#endif
+
+unsigned long arch_scale_smt_power(struct sched_domain *sd, int cpu)
+{
+ int sibling;
+ int idle_count = 0;
+ int thread;
+
+ struct cpumask *sibling_map = sched_domain_span(sd);
+
+ unsigned long weight = cpumask_weight(sibling_map);
+ unsigned long smt_gain = sd->smt_gain;
+
+ if (cpu_has_feature(CPU_FTR_ASYNC_SMT4) && weight == 4) {
+ for_each_cpu(sibling, sibling_map) {
+ if (idle_cpu(sibling))
+ idle_count++;
+ }
+
+ /* the following section attempts to tweak cpu power based
+ * on current idleness of the threads dynamically at runtime
+ */
+ if (idle_count > 1) {
+ thread = cpu_thread_in_core(cpu);
+ if (thread < 2) {
+ /* add 75 % to thread power */
+ smt_gain += (smt_gain >> 1) + (smt_gain >> 2);
+ } else {
+ /* subtract 75 % to thread power */
+ smt_gain = smt_gain >> 2;
+ }
+ }
+ }
+
+ /* default smt gain is 1178, weight is # of SMT threads */
+ switch (weight) {
+ case 1:
+ /*divide by 1, do nothing*/
+ break;
+ case 2:
+ smt_gain = smt_gain >> 1;
+ break;
+ case 4:
+ smt_gain = smt_gain >> 2;
+ break;
+ default:
+ smt_gain /= weight;
+ break;
+ }
+
+ return smt_gain;
+
+}
Index: linux-2.6.git/arch/powerpc/include/asm/cputable.h
===================================================================
--- linux-2.6.git.orig/arch/powerpc/include/asm/cputable.h
+++ linux-2.6.git/arch/powerpc/include/asm/cputable.h
@@ -195,6 +195,7 @@ extern const char *powerpc_base_platform
#define CPU_FTR_SAO LONG_ASM_CONST(0x0020000000000000)
#define CPU_FTR_CP_USE_DCBTZ LONG_ASM_CONST(0x0040000000000000)
#define CPU_FTR_UNALIGNED_LD_STD LONG_ASM_CONST(0x0080000000000000)
+#define CPU_FTR_ASYNC_SMT4 LONG_ASM_CONST(0x0100000000000000)
#ifndef __ASSEMBLY__
@@ -409,7 +410,7 @@ extern const char *powerpc_base_platform
CPU_FTR_MMCRA | CPU_FTR_SMT | \
CPU_FTR_COHERENT_ICACHE | CPU_FTR_LOCKLESS_TLBIE | \
CPU_FTR_PURR | CPU_FTR_SPURR | CPU_FTR_REAL_LE | \
- CPU_FTR_DSCR | CPU_FTR_SAO)
+ CPU_FTR_DSCR | CPU_FTR_SAO | CPU_FTR_ASYNC_SMT4)
#define CPU_FTRS_CELL (CPU_FTR_USE_TB | CPU_FTR_LWSYNC | \
CPU_FTR_PPCAS_ARCH_V2 | CPU_FTR_CTRL | \
CPU_FTR_ALTIVEC_COMP | CPU_FTR_MMCRA | CPU_FTR_SMT | \
On Tue, 2010-01-26 at 17:28 -0600, Joel Schopp wrote:
> On Power7 processors running in SMT4 mode with 2, 3, or 4 idle threads
> there is performance benefit to idling the higher numbered threads in
> the core.
>
> This patch implements arch_scale_smt_power to dynamically update smt
> thread power in these idle cases in order to prefer threads 0,1 over
> threads 2,3 within a core.
>
> v2 - Same functionality as v1, better coding style.
Better. Some more comments...
> Signed-off-by: Joel Schopp <[email protected]>
> ---
> Version 2 addresses style and optimization, same basic functionality
> Index: linux-2.6.git/arch/powerpc/kernel/smp.c
> ===================================================================
> --- linux-2.6.git.orig/arch/powerpc/kernel/smp.c
> +++ linux-2.6.git/arch/powerpc/kernel/smp.c
> @@ -620,3 +620,55 @@ void __cpu_die(unsigned int cpu)
> smp_ops->cpu_die(cpu);
> }
> #endif
> +
> +unsigned long arch_scale_smt_power(struct sched_domain *sd, int cpu)
> +{
> + int sibling;
> + int idle_count = 0;
> + int thread;
> +
> + struct cpumask *sibling_map = sched_domain_span(sd);
What about an early exit if !cpu_has_feature(CPU_FTR_SMT) ? That would
de-facto compile it out for 32-bit CPU platforms that don't support SMT
at all and avoid some overhead on POWER3,4,970...
> + unsigned long weight = cpumask_weight(sibling_map);
> + unsigned long smt_gain = sd->smt_gain;
> +
> + if (cpu_has_feature(CPU_FTR_ASYNC_SMT4) && weight == 4) {
So that will only handle the case where all 4 threads are online right ?
There is no provision for the case where the user play tricks like
offlining thread, in which case it will stop trying to "push down"
processes right ? Not a big deal per-se I suppose, just something to be
aware of.
Also, can you add a comment as to why this is done in the code itself ?
above the if (cpu_has_feature(...)) statement.
> + for_each_cpu(sibling, sibling_map) {
> + if (idle_cpu(sibling))
> + idle_count++;
> + }
> +
> + /* the following section attempts to tweak cpu power based
> + * on current idleness of the threads dynamically at runtime
> + */
> + if (idle_count > 1) {
> + thread = cpu_thread_in_core(cpu);
> + if (thread < 2) {
> + /* add 75 % to thread power */
> + smt_gain += (smt_gain >> 1) + (smt_gain >> 2);
> + } else {
> + /* subtract 75 % to thread power */
> + smt_gain = smt_gain >> 2;
> + }
> + }
> + }
> +
> + /* default smt gain is 1178, weight is # of SMT threads */
> + switch (weight) {
> + case 1:
> + /*divide by 1, do nothing*/
> + break;
> + case 2:
> + smt_gain = smt_gain >> 1;
> + break;
> + case 4:
> + smt_gain = smt_gain >> 2;
> + break;
> + default:
> + smt_gain /= weight;
> + break;
> + }
> +
> + return smt_gain;
> +}
Appart from that, it looks allright to me.
Cheers,
Ben.
>
> What about an early exit if !cpu_has_feature(CPU_FTR_SMT) ? That would
> de-facto compile it out for 32-bit CPU platforms that don't support SMT
> at all and avoid some overhead on POWER3,4,970...
>
If the SD_SHARE_CPUPOWER flag isn't set for the sched domain this
function isn't called. So an extra check here is wasteful.
>
>> + unsigned long weight = cpumask_weight(sibling_map);
>> + unsigned long smt_gain = sd->smt_gain;
>> +
>> + if (cpu_has_feature(CPU_FTR_ASYNC_SMT4) && weight == 4) {
>>
>
> So that will only handle the case where all 4 threads are online right ?
> There is no provision for the case where the user play tricks like
> offlining thread, in which case it will stop trying to "push down"
> processes right ? Not a big deal per-se I suppose, just something to be
> aware of.
>
I've tested it with manually offlined threads and it behaves as I'd like
it to.
> Also, can you add a comment as to why this is done in the code itself ?
> above the if (cpu_has_feature(...)) statement.
>
OK. v3 coming soon with the comment.
The new Power7 processor has 4 way SMT. This 4 way SMT benefits from
dynamic power updates that arch_scale_smt_power was designed to provide.
The first patch fixes a generic scheduler bug necessary for arch_scale_smt
to properly function. The second patch implements arch_scale_smt_power
for powerpc, and in particular for Power7 processors.
Version 2 changes:
- Drop Patch 1 from the original series since it's in the -tip tree now
- Move enabling the cpu feature into it's own patch (now patch 1)
- Add stubbing out broken x86 implementation to patch 1
- clean up coding style in patch 2
Version 3 changes:
- Added a comment to Patch 2
---
Joel Schopp (2):
powerpc: implement arch_scale_smt_power for Power7
sched: enable ARCH_POWER
arch/powerpc/include/asm/cputable.h | 3 +
arch/powerpc/kernel/smp.c | 56
++++++++++++++++++++++++++++++++++++
arch/x86/kernel/cpu/sched.c | 6 +--
kernel/sched_features.h | 2 -
4 files changed, 61 insertions(+), 6 deletions(-)
Enable the scheduler feature that allows use of arch_scale_smt_power. Stub out
the broken x86 implementation.
Signed-off-by: Joel Schopp <[email protected]>
---
Index: linux-2.6.git/kernel/sched_features.h
===================================================================
--- linux-2.6.git.orig/kernel/sched_features.h
+++ linux-2.6.git/kernel/sched_features.h
@@ -102,7 +102,7 @@ SCHED_FEAT(CACHE_HOT_BUDDY, 1)
/*
* Use arch dependent cpu power functions
*/
-SCHED_FEAT(ARCH_POWER, 0)
+SCHED_FEAT(ARCH_POWER, 1)
SCHED_FEAT(HRTICK, 0)
SCHED_FEAT(DOUBLE_TICK, 0)
Index: linux-2.6.git/arch/x86/kernel/cpu/sched.c
===================================================================
--- linux-2.6.git.orig/arch/x86/kernel/cpu/sched.c
+++ linux-2.6.git/arch/x86/kernel/cpu/sched.c
@@ -44,11 +44,9 @@ unsigned long arch_scale_freq_power(stru
unsigned long arch_scale_smt_power(struct sched_domain *sd, int cpu)
{
/*
- * aperf/mperf already includes the smt gain
+ * aperf/mperf already includes the smt gain, but represents capacity
+ * as 0 when idle. So for now just return default.
*/
- if (boot_cpu_has(X86_FEATURE_APERFMPERF))
- return SCHED_LOAD_SCALE;
-
return default_scale_smt_power(sd, cpu);
}
On Power7 processors running in SMT4 mode with 2, 3, or 4 idle threads
there is performance benefit to idling the higher numbered threads in
the core.
This patch implements arch_scale_smt_power to dynamically update smt
thread power in these idle cases in order to prefer threads 0,1 over
threads 2,3 within a core.
Signed-off-by: Joel Schopp <[email protected]>
---
Version 2 addresses style and optimization, same basic functionality
Version 3 adds a comment
On Power7 processors running in SMT4 mode with 2, 3, or 4 idle threads
there is performance benefit to idling the higher numbered threads in
the core.
This patch implements arch_scale_smt_power to dynamically update smt
thread power in these idle cases in order to prefer threads 0,1 over
threads 2,3 within a core.
Signed-off-by: Joel Schopp <[email protected]>
---
Version 2 addresses style and optimization, same basic functionality
Index: linux-2.6.git/arch/powerpc/kernel/smp.c
===================================================================
--- linux-2.6.git.orig/arch/powerpc/kernel/smp.c
+++ linux-2.6.git/arch/powerpc/kernel/smp.c
@@ -620,3 +620,59 @@ void __cpu_die(unsigned int cpu)
smp_ops->cpu_die(cpu);
}
#endif
+
+unsigned long arch_scale_smt_power(struct sched_domain *sd, int cpu)
+{
+ int sibling;
+ int idle_count = 0;
+ int thread;
+
+ /* Setup the default weight and smt_gain used by most cpus for SMT
+ * Power. Doing this right away covers the default case and can be
+ * used by cpus that modify it dynamically.
+ */
+ struct cpumask *sibling_map = sched_domain_span(sd);
+ unsigned long weight = cpumask_weight(sibling_map);
+ unsigned long smt_gain = sd->smt_gain;
+
+
+ if (cpu_has_feature(CPU_FTR_ASYNC_SMT4) && weight == 4) {
+ for_each_cpu(sibling, sibling_map) {
+ if (idle_cpu(sibling))
+ idle_count++;
+ }
+
+ /* the following section attempts to tweak cpu power based
+ * on current idleness of the threads dynamically at runtime
+ */
+ if (idle_count > 1) {
+ thread = cpu_thread_in_core(cpu);
+ if (thread < 2) {
+ /* add 75 % to thread power */
+ smt_gain += (smt_gain >> 1) + (smt_gain >> 2);
+ } else {
+ /* subtract 75 % to thread power */
+ smt_gain = smt_gain >> 2;
+ }
+ }
+ }
+
+ /* default smt gain is 1178, weight is # of SMT threads */
+ switch (weight) {
+ case 1:
+ /*divide by 1, do nothing*/
+ break;
+ case 2:
+ smt_gain = smt_gain >> 1;
+ break;
+ case 4:
+ smt_gain = smt_gain >> 2;
+ break;
+ default:
+ smt_gain /= weight;
+ break;
+ }
+
+ return smt_gain;
+
+}
Index: linux-2.6.git/arch/powerpc/include/asm/cputable.h
===================================================================
--- linux-2.6.git.orig/arch/powerpc/include/asm/cputable.h
+++ linux-2.6.git/arch/powerpc/include/asm/cputable.h
@@ -195,6 +195,7 @@ extern const char *powerpc_base_platform
#define CPU_FTR_SAO LONG_ASM_CONST(0x0020000000000000)
#define CPU_FTR_CP_USE_DCBTZ LONG_ASM_CONST(0x0040000000000000)
#define CPU_FTR_UNALIGNED_LD_STD LONG_ASM_CONST(0x0080000000000000)
+#define CPU_FTR_ASYNC_SMT4 LONG_ASM_CONST(0x0100000000000000)
#ifndef __ASSEMBLY__
@@ -409,7 +410,7 @@ extern const char *powerpc_base_platform
CPU_FTR_MMCRA | CPU_FTR_SMT | \
CPU_FTR_COHERENT_ICACHE | CPU_FTR_LOCKLESS_TLBIE | \
CPU_FTR_PURR | CPU_FTR_SPURR | CPU_FTR_REAL_LE | \
- CPU_FTR_DSCR | CPU_FTR_SAO)
+ CPU_FTR_DSCR | CPU_FTR_SAO | CPU_FTR_ASYNC_SMT4)
#define CPU_FTRS_CELL (CPU_FTR_USE_TB | CPU_FTR_LWSYNC | \
CPU_FTR_PPCAS_ARCH_V2 | CPU_FTR_CTRL | \
CPU_FTR_ALTIVEC_COMP | CPU_FTR_MMCRA | CPU_FTR_SMT | \
On Power7 processors running in SMT4 mode with 2, 3, or 4 idle threads
there is performance benefit to idling the higher numbered threads in
the core.
This patch implements arch_scale_smt_power to dynamically update smt
thread power in these idle cases in order to prefer threads 0,1 over
threads 2,3 within a core.
Signed-off-by: Joel Schopp <[email protected]>
---
Version 2 addresses style and optimization, same basic functionality
Version 3 adds a comment, resent due to mailing format error
Index: linux-2.6.git/arch/powerpc/kernel/smp.c
===================================================================
--- linux-2.6.git.orig/arch/powerpc/kernel/smp.c
+++ linux-2.6.git/arch/powerpc/kernel/smp.c
@@ -620,3 +620,59 @@ void __cpu_die(unsigned int cpu)
smp_ops->cpu_die(cpu);
}
#endif
+
+unsigned long arch_scale_smt_power(struct sched_domain *sd, int cpu)
+{
+ int sibling;
+ int idle_count = 0;
+ int thread;
+
+ /* Setup the default weight and smt_gain used by most cpus for SMT
+ * Power. Doing this right away covers the default case and can be
+ * used by cpus that modify it dynamically.
+ */
+ struct cpumask *sibling_map = sched_domain_span(sd);
+ unsigned long weight = cpumask_weight(sibling_map);
+ unsigned long smt_gain = sd->smt_gain;
+
+
+ if (cpu_has_feature(CPU_FTR_ASYNC_SMT4) && weight == 4) {
+ for_each_cpu(sibling, sibling_map) {
+ if (idle_cpu(sibling))
+ idle_count++;
+ }
+
+ /* the following section attempts to tweak cpu power based
+ * on current idleness of the threads dynamically at runtime
+ */
+ if (idle_count > 1) {
+ thread = cpu_thread_in_core(cpu);
+ if (thread < 2) {
+ /* add 75 % to thread power */
+ smt_gain += (smt_gain >> 1) + (smt_gain >> 2);
+ } else {
+ /* subtract 75 % to thread power */
+ smt_gain = smt_gain >> 2;
+ }
+ }
+ }
+
+ /* default smt gain is 1178, weight is # of SMT threads */
+ switch (weight) {
+ case 1:
+ /*divide by 1, do nothing*/
+ break;
+ case 2:
+ smt_gain = smt_gain >> 1;
+ break;
+ case 4:
+ smt_gain = smt_gain >> 2;
+ break;
+ default:
+ smt_gain /= weight;
+ break;
+ }
+
+ return smt_gain;
+
+}
Index: linux-2.6.git/arch/powerpc/include/asm/cputable.h
===================================================================
--- linux-2.6.git.orig/arch/powerpc/include/asm/cputable.h
+++ linux-2.6.git/arch/powerpc/include/asm/cputable.h
@@ -195,6 +195,7 @@ extern const char *powerpc_base_platform
#define CPU_FTR_SAO LONG_ASM_CONST(0x0020000000000000)
#define CPU_FTR_CP_USE_DCBTZ LONG_ASM_CONST(0x0040000000000000)
#define CPU_FTR_UNALIGNED_LD_STD LONG_ASM_CONST(0x0080000000000000)
+#define CPU_FTR_ASYNC_SMT4 LONG_ASM_CONST(0x0100000000000000)
#ifndef __ASSEMBLY__
@@ -409,7 +410,7 @@ extern const char *powerpc_base_platform
CPU_FTR_MMCRA | CPU_FTR_SMT | \
CPU_FTR_COHERENT_ICACHE | CPU_FTR_LOCKLESS_TLBIE | \
CPU_FTR_PURR | CPU_FTR_SPURR | CPU_FTR_REAL_LE | \
- CPU_FTR_DSCR | CPU_FTR_SAO)
+ CPU_FTR_DSCR | CPU_FTR_SAO | CPU_FTR_ASYNC_SMT4)
#define CPU_FTRS_CELL (CPU_FTR_USE_TB | CPU_FTR_LWSYNC | \
CPU_FTR_PPCAS_ARCH_V2 | CPU_FTR_CTRL | \
CPU_FTR_ALTIVEC_COMP | CPU_FTR_MMCRA | CPU_FTR_SMT | \
> I've tested it with manually offlined threads and it behaves as I'd like
> it to.
Which is ? IE. I'm happy that you like how it behaves, but I'd like to u
understand how that is so I can make sure I'm also happy with it :-)
Cheers,
Ben.
On Thu, 2010-01-28 at 17:24 -0600, Joel Schopp wrote:
> On Power7 processors running in SMT4 mode with 2, 3, or 4 idle threads
> there is performance benefit to idling the higher numbered threads in
> the core.
>
> This patch implements arch_scale_smt_power to dynamically update smt
> thread power in these idle cases in order to prefer threads 0,1 over
> threads 2,3 within a core.
Almost there :-) Joel, Peter, can you help me figure something out tho ?
On machine that don't have SMT, I would like to avoid calling
arch_scale_smt_power() at all if possible (in addition to not compiling
it in if SMT is not enabled in .config).
Now, I must say I'm utterly confused by how the domains are setup and I
haven't quite managed to sort it out... it looks to me that
SD_SHARE_CPUPOWER is always going to be set on all CPUs when the config
option is set (though each CPU will have its own domain) or am I
misguided ? IE. Is there any sense in having at least a fast exit path
out of arch_scale_smt_power() for non-SMT CPUs ?
Joel, can you look at compiling it out when SMT is not set ? We don't
want to bloat SMP kernels for 32-bit non-SMT embedded platforms.
Oh, and one minor nit:
> Signed-off-by: Joel Schopp <[email protected]>
> ---
> Version 2 addresses style and optimization, same basic functionality
> Version 3 adds a comment, resent due to mailing format error
> Index: linux-2.6.git/arch/powerpc/kernel/smp.c
> ===================================================================
> --- linux-2.6.git.orig/arch/powerpc/kernel/smp.c
> +++ linux-2.6.git/arch/powerpc/kernel/smp.c
> @@ -620,3 +620,59 @@ void __cpu_die(unsigned int cpu)
> smp_ops->cpu_die(cpu);
> }
> #endif
^^^ Please add the /* CONFIG_CPU_HOTPLUG */ (or whatever it is) that's
missing after that #endif :-)
Cheers,
Ben.
On Fri, 2010-01-29 at 12:23 +1100, Benjamin Herrenschmidt wrote:
> On machine that don't have SMT, I would like to avoid calling
> arch_scale_smt_power() at all if possible (in addition to not compiling
> it in if SMT is not enabled in .config).
>
> Now, I must say I'm utterly confused by how the domains are setup and I
> haven't quite managed to sort it out... it looks to me that
> SD_SHARE_CPUPOWER is always going to be set on all CPUs when the config
> option is set (though each CPU will have its own domain) or am I
> misguided ? IE. Is there any sense in having at least a fast exit path
> out of arch_scale_smt_power() for non-SMT CPUs ?
The sched_domain creation code is a f'ing stink pile that hurts
everybody's brain.
The AMD magny-cours people sort of cleaned it up a bit but didn't go
nearly far enough. Doing so is somewhere on my todo list, but sadly that
thing is way larger than my spare time.
Now SD_SHARE_CPUPOWER _should_ only be set for SMT domains, because only
SMT siblings share cpupower.
SD_SHARE_PKG_RESOURCES _should_ be set for both SMT and MC, because they
all share the same cache domain.
If it all works out that way in practice on powerpc is another question
entirely ;-)
That said, I'm still not entirely convinced I like this usage of
cpupower, its supposed to be a normalization scale for load-balancing,
not a placement hook.
I'd be much happier with a SD_GROUP_ORDER or something like that, that
works together with SD_PREFER_SIBLING to pack active tasks to cpus in
ascending group order.
On Thu, Jan 28, 2010 at 05:20:55PM -0600, Joel Schopp wrote:
> On Power7 processors running in SMT4 mode with 2, 3, or 4 idle threads
> there is performance benefit to idling the higher numbered threads in
> the core.
>
Really 2, 3, or 4? When you have 4 idle threads out of 4, performance
becomes a minor concern, no? ;-)
Gabriel
Gabriel Paubert wrote:
> On Thu, Jan 28, 2010 at 05:20:55PM -0600, Joel Schopp wrote:
>
>> On Power7 processors running in SMT4 mode with 2, 3, or 4 idle threads
>> there is performance benefit to idling the higher numbered threads in
>> the core.
>>
>>
>
> Really 2, 3, or 4? When you have 4 idle threads out of 4, performance
> becomes a minor concern, no? ;-)
>
> Gabriel
>
Yes, but going from 4 idle to 3 idle you want to keep the slanted
weights. If you ignored 4 you'd place wrong and then correct it after
the fact.
> That said, I'm still not entirely convinced I like this usage of
> cpupower, its supposed to be a normalization scale for load-balancing,
> not a placement hook.
>
Even if you do a placement hook you'll need to address it in the load
balancing as well. Consider a single 4 thread SMT core with 4 running
tasks. If 2 of them exit the remaining 2 will need to be load balanced
within the core in a way that takes into account the dynamic nature of
the thread power. This patch does that.
> I'd be much happier with a SD_GROUP_ORDER or something like that, that
> works together with SD_PREFER_SIBLING to pack active tasks to cpus in
> ascending group order.
>
>
I don't see this load-balancing patch as mutually exclusive with a patch
to fix placement. But even if it is a mutually exclusive solution there
is no reason we can't fix things now with this patch and then later take
it out when it's fixed another way. This patch series is
straightforward, non-intrusive, and without it the scheduler is broken
on this processor.
Benjamin Herrenschmidt wrote:
> On Thu, 2010-01-28 at 17:24 -0600, Joel Schopp wrote:
>
>> On Power7 processors running in SMT4 mode with 2, 3, or 4 idle threads
>> there is performance benefit to idling the higher numbered threads in
>> the core.
>>
>> This patch implements arch_scale_smt_power to dynamically update smt
>> thread power in these idle cases in order to prefer threads 0,1 over
>> threads 2,3 within a core.
>>
>
> Almost there :-) Joel, Peter, can you help me figure something out tho ?
>
> On machine that don't have SMT, I would like to avoid calling
> arch_scale_smt_power() at all if possible (in addition to not compiling
> it in if SMT is not enabled in .config).
>
> Now, I must say I'm utterly confused by how the domains are setup and I
> haven't quite managed to sort it out... it looks to me that
> SD_SHARE_CPUPOWER is always going to be set on all CPUs when the config
> option is set (though each CPU will have its own domain) or am I
> misguided ? IE. Is there any sense in having at least a fast exit path
> out of arch_scale_smt_power() for non-SMT CPUs ?
>
> Joel, can you look at compiling it out when SMT is not set ? We don't
> want to bloat SMP kernels for 32-bit non-SMT embedded platforms.
>
I can wrap the powerpc definition of arch_scale_smt in an #ifdef, if
it's not there the scheduler uses the default, which is the same as it
uses if SMT isn't compiled.
The new Power7 processor has 4 way SMT. This 4 way SMT benefits from
dynamic power updates that arch_scale_smt_power was designed to provide.
The first patch fixes a generic scheduler bug necessary for arch_scale_smt
to properly function. The second patch implements arch_scale_smt_power
for powerpc, and in particular for Power7 processors.
Version 2 changes:
- Drop Patch 1 from the original series since it's in the -tip tree now
- Move enabling the cpu feature into it's own patch (now patch 1)
- Add stubbing out broken x86 implementation to patch 1
- clean up coding style in patch 2
Version 3 changes:
- Added a comment to Patch 2
Version 4 changes:
- Wrap arch_scale_smt_power() in #ifdef CONFIG_SCHED_SMT in Patch 2
---
Joel Schopp (2):
sched: enable ARCH_POWER
powerpc: implement arch_scale_smt_power for Power7
arch/powerpc/include/asm/cputable.h | 3 +
arch/powerpc/kernel/smp.c | 58
++++++++++++++++++++++++++++++++++++
arch/x86/kernel/cpu/sched.c | 6 +--
kernel/sched_features.h | 2 -
4 files changed, 63 insertions(+), 6 deletions(-)
Enable the scheduler feature that allows use of arch_scale_smt_power. Stub out
the broken x86 implementation.
Signed-off-by: Joel Schopp <[email protected]>
---
Index: linux-2.6.git/kernel/sched_features.h
===================================================================
--- linux-2.6.git.orig/kernel/sched_features.h
+++ linux-2.6.git/kernel/sched_features.h
@@ -102,7 +102,7 @@ SCHED_FEAT(CACHE_HOT_BUDDY, 1)
/*
* Use arch dependent cpu power functions
*/
-SCHED_FEAT(ARCH_POWER, 0)
+SCHED_FEAT(ARCH_POWER, 1)
SCHED_FEAT(HRTICK, 0)
SCHED_FEAT(DOUBLE_TICK, 0)
Index: linux-2.6.git/arch/x86/kernel/cpu/sched.c
===================================================================
--- linux-2.6.git.orig/arch/x86/kernel/cpu/sched.c
+++ linux-2.6.git/arch/x86/kernel/cpu/sched.c
@@ -44,11 +44,9 @@ unsigned long arch_scale_freq_power(stru
unsigned long arch_scale_smt_power(struct sched_domain *sd, int cpu)
{
/*
- * aperf/mperf already includes the smt gain
+ * aperf/mperf already includes the smt gain, but represents capacity
+ * as 0 when idle. So for now just return default.
*/
- if (boot_cpu_has(X86_FEATURE_APERFMPERF))
- return SCHED_LOAD_SCALE;
-
return default_scale_smt_power(sd, cpu);
}
On Power7 processors running in SMT4 mode with 2, 3, or 4 idle threads
there is performance benefit to idling the higher numbered threads in
the core.
This patch implements arch_scale_smt_power to dynamically update smt
thread power in these idle cases in order to prefer threads 0,1 over
threads 2,3 within a core.
Signed-off-by: Joel Schopp <[email protected]>
---
Version 3 adds the #ifdef to avoid compiling on kernels that don't need it
Index: linux-2.6.git/arch/powerpc/kernel/smp.c
===================================================================
--- linux-2.6.git.orig/arch/powerpc/kernel/smp.c
+++ linux-2.6.git/arch/powerpc/kernel/smp.c
@@ -620,3 +620,61 @@ void __cpu_die(unsigned int cpu)
smp_ops->cpu_die(cpu);
}
#endif
+
+#ifdef CONFIG_SCHED_SMT
+unsigned long arch_scale_smt_power(struct sched_domain *sd, int cpu)
+{
+ int sibling;
+ int idle_count = 0;
+ int thread;
+
+ /* Setup the default weight and smt_gain used by most cpus for SMT
+ * Power. Doing this right away covers the default case and can be
+ * used by cpus that modify it dynamically.
+ */
+ struct cpumask *sibling_map = sched_domain_span(sd);
+ unsigned long weight = cpumask_weight(sibling_map);
+ unsigned long smt_gain = sd->smt_gain;
+
+
+ if (cpu_has_feature(CPU_FTR_ASYNC_SMT4) && weight == 4) {
+ for_each_cpu(sibling, sibling_map) {
+ if (idle_cpu(sibling))
+ idle_count++;
+ }
+
+ /* the following section attempts to tweak cpu power based
+ * on current idleness of the threads dynamically at runtime
+ */
+ if (idle_count > 1) {
+ thread = cpu_thread_in_core(cpu);
+ if (thread < 2) {
+ /* add 75 % to thread power */
+ smt_gain += (smt_gain >> 1) + (smt_gain >> 2);
+ } else {
+ /* subtract 75 % to thread power */
+ smt_gain = smt_gain >> 2;
+ }
+ }
+ }
+
+ /* default smt gain is 1178, weight is # of SMT threads */
+ switch (weight) {
+ case 1:
+ /*divide by 1, do nothing*/
+ break;
+ case 2:
+ smt_gain = smt_gain >> 1;
+ break;
+ case 4:
+ smt_gain = smt_gain >> 2;
+ break;
+ default:
+ smt_gain /= weight;
+ break;
+ }
+
+ return smt_gain;
+
+}
+#endif
Index: linux-2.6.git/arch/powerpc/include/asm/cputable.h
===================================================================
--- linux-2.6.git.orig/arch/powerpc/include/asm/cputable.h
+++ linux-2.6.git/arch/powerpc/include/asm/cputable.h
@@ -195,6 +195,7 @@ extern const char *powerpc_base_platform
#define CPU_FTR_SAO LONG_ASM_CONST(0x0020000000000000)
#define CPU_FTR_CP_USE_DCBTZ LONG_ASM_CONST(0x0040000000000000)
#define CPU_FTR_UNALIGNED_LD_STD LONG_ASM_CONST(0x0080000000000000)
+#define CPU_FTR_ASYNC_SMT4 LONG_ASM_CONST(0x0100000000000000)
#ifndef __ASSEMBLY__
@@ -409,7 +410,7 @@ extern const char *powerpc_base_platform
CPU_FTR_MMCRA | CPU_FTR_SMT | \
CPU_FTR_COHERENT_ICACHE | CPU_FTR_LOCKLESS_TLBIE | \
CPU_FTR_PURR | CPU_FTR_SPURR | CPU_FTR_REAL_LE | \
- CPU_FTR_DSCR | CPU_FTR_SAO)
+ CPU_FTR_DSCR | CPU_FTR_SAO | CPU_FTR_ASYNC_SMT4)
#define CPU_FTRS_CELL (CPU_FTR_USE_TB | CPU_FTR_LWSYNC | \
CPU_FTR_PPCAS_ARCH_V2 | CPU_FTR_CTRL | \
CPU_FTR_ALTIVEC_COMP | CPU_FTR_MMCRA | CPU_FTR_SMT | \
On Fri, 2010-02-05 at 14:57 -0600, Joel Schopp wrote:
> On Power7 processors running in SMT4 mode with 2, 3, or 4 idle threads
> there is performance benefit to idling the higher numbered threads in
> the core.
>
> This patch implements arch_scale_smt_power to dynamically update smt
> thread power in these idle cases in order to prefer threads 0,1 over
> threads 2,3 within a core.
>
> Signed-off-by: Joel Schopp <[email protected]>
> ---
> Index: linux-2.6.git/arch/powerpc/kernel/smp.c
> ===================================================================
> --- linux-2.6.git.orig/arch/powerpc/kernel/smp.c
> +++ linux-2.6.git/arch/powerpc/kernel/smp.c
> @@ -620,3 +620,61 @@ void __cpu_die(unsigned int cpu)
> smp_ops->cpu_die(cpu);
> }
> #endif
> +
> +#ifdef CONFIG_SCHED_SMT
> +unsigned long arch_scale_smt_power(struct sched_domain *sd, int cpu)
> +{
> + int sibling;
> + int idle_count = 0;
> + int thread;
> +
> + /* Setup the default weight and smt_gain used by most cpus for SMT
> + * Power. Doing this right away covers the default case and can be
> + * used by cpus that modify it dynamically.
> + */
> + struct cpumask *sibling_map = sched_domain_span(sd);
> + unsigned long weight = cpumask_weight(sibling_map);
> + unsigned long smt_gain = sd->smt_gain;
> +
> +
> + if (cpu_has_feature(CPU_FTR_ASYNC_SMT4) && weight == 4) {
> + for_each_cpu(sibling, sibling_map) {
> + if (idle_cpu(sibling))
> + idle_count++;
> + }
> +
> + /* the following section attempts to tweak cpu power based
> + * on current idleness of the threads dynamically at runtime
> + */
> + if (idle_count > 1) {
> + thread = cpu_thread_in_core(cpu);
> + if (thread < 2) {
> + /* add 75 % to thread power */
> + smt_gain += (smt_gain >> 1) + (smt_gain >> 2);
> + } else {
> + /* subtract 75 % to thread power */
> + smt_gain = smt_gain >> 2;
> + }
> + }
> + }
> +
> + /* default smt gain is 1178, weight is # of SMT threads */
> + switch (weight) {
> + case 1:
> + /*divide by 1, do nothing*/
> + break;
> + case 2:
> + smt_gain = smt_gain >> 1;
> + break;
> + case 4:
> + smt_gain = smt_gain >> 2;
> + break;
> + default:
> + smt_gain /= weight;
> + break;
> + }
> +
> + return smt_gain;
> +
> +}
> +#endif
Suppose for a moment we have 2 threads (hot-unplugged thread 1 and 3, we
can construct an equivalent but more complex example for 4 threads), and
we have 4 tasks, 3 SCHED_OTHER of equal nice level and 1 SCHED_FIFO, the
SCHED_FIFO task will consume exactly 50% walltime of whatever cpu it
ends up on.
In that situation, provided that each cpu's cpu_power is of equal
measure, scale_rt_power() ensures that we run 2 SCHED_OTHER tasks on the
cpu that doesn't run the RT task, and 1 SCHED_OTHER task next to the RT
task, so that each task consumes 50%, which is all fair and proper.
However, if you do the above, thread 0 will have +75% = 1.75 and thread
2 will have -75% = 0.25, then if the RT task will land on thread 0,
we'll be having: 0.875 vs 0.25, or on thread 3, 1.75 vs 0.125. In either
case thread 0 will receive too many (if not all) SCHED_OTHER tasks.
That is, unless these threads 2 and 3 really are _that_ weak, at which
point one wonders why IBM bothered with the silicon ;-)
So tell me again, why is fiddling with the cpu_power a good placement
tool?
> Suppose for a moment we have 2 threads (hot-unplugged thread 1 and 3, we
> can construct an equivalent but more complex example for 4 threads), and
> we have 4 tasks, 3 SCHED_OTHER of equal nice level and 1 SCHED_FIFO, the
> SCHED_FIFO task will consume exactly 50% walltime of whatever cpu it
> ends up on.
>
> In that situation, provided that each cpu's cpu_power is of equal
> measure, scale_rt_power() ensures that we run 2 SCHED_OTHER tasks on the
> cpu that doesn't run the RT task, and 1 SCHED_OTHER task next to the RT
> task, so that each task consumes 50%, which is all fair and proper.
>
> However, if you do the above, thread 0 will have +75% = 1.75 and thread
> 2 will have -75% = 0.25, then if the RT task will land on thread 0,
> we'll be having: 0.875 vs 0.25, or on thread 3, 1.75 vs 0.125. In either
> case thread 0 will receive too many (if not all) SCHED_OTHER tasks.
>
> That is, unless these threads 2 and 3 really are _that_ weak, at which
> point one wonders why IBM bothered with the silicon ;-)
Peter,
2 & 3 aren't weaker than 0 & 1 but....
The core has dynamic SMT mode switching which is controlled by the
hypervisor (IBM's PHYP). There are 3 SMT modes:
SMT1 uses thread 0
SMT2 uses threads 0 & 1
SMT4 uses threads 0, 1, 2 & 3
When in any particular SMT mode, all threads have the same performance
as each other (ie. at any moment in time, all threads perform the same).
The SMT mode switching works such that when linux has threads 2 & 3 idle
and 0 & 1 active, it will cede (H_CEDE hypercall) threads 2 and 3 in the
idle loop and the hypervisor will automatically switch to SMT2 for that
core (independent of other cores). The opposite is not true, so if
threads 0 & 1 are idle and 2 & 3 are active, we will stay in SMT4 mode.
Similarly if thread 0 is active and threads 1, 2 & 3 are idle, we'll go
into SMT1 mode.
If we can get the core into a lower SMT mode (SMT1 is best), the threads
will perform better (since they share less core resources). Hence when
we have idle threads, we want them to be the higher ones.
So to answer your question, threads 2 and 3 aren't weaker than the other
threads when in SMT4 mode. It's that if we idle threads 2 & 3, threads
0 & 1 will speed up since we'll move to SMT2 mode.
I'm pretty vague on linux scheduler details, so I'm a bit at sea as to
how to solve this. Can you suggest any mechanisms we currently have in
the kernel to reflect these properties, or do you think we need to
develop something new? If so, any pointers as to where we should look?
Thanks,
Mikey
On Thu, 2010-02-18 at 09:20 +1100, Michael Neuling wrote:
> > Suppose for a moment we have 2 threads (hot-unplugged thread 1 and 3, we
> > can construct an equivalent but more complex example for 4 threads), and
> > we have 4 tasks, 3 SCHED_OTHER of equal nice level and 1 SCHED_FIFO, the
> > SCHED_FIFO task will consume exactly 50% walltime of whatever cpu it
> > ends up on.
> >
> > In that situation, provided that each cpu's cpu_power is of equal
> > measure, scale_rt_power() ensures that we run 2 SCHED_OTHER tasks on the
> > cpu that doesn't run the RT task, and 1 SCHED_OTHER task next to the RT
> > task, so that each task consumes 50%, which is all fair and proper.
> >
> > However, if you do the above, thread 0 will have +75% = 1.75 and thread
> > 2 will have -75% = 0.25, then if the RT task will land on thread 0,
> > we'll be having: 0.875 vs 0.25, or on thread 3, 1.75 vs 0.125. In either
> > case thread 0 will receive too many (if not all) SCHED_OTHER tasks.
> >
> > That is, unless these threads 2 and 3 really are _that_ weak, at which
> > point one wonders why IBM bothered with the silicon ;-)
>
> Peter,
>
> 2 & 3 aren't weaker than 0 & 1 but....
>
> The core has dynamic SMT mode switching which is controlled by the
> hypervisor (IBM's PHYP). There are 3 SMT modes:
> SMT1 uses thread 0
> SMT2 uses threads 0 & 1
> SMT4 uses threads 0, 1, 2 & 3
> When in any particular SMT mode, all threads have the same performance
> as each other (ie. at any moment in time, all threads perform the same).
>
> The SMT mode switching works such that when linux has threads 2 & 3 idle
> and 0 & 1 active, it will cede (H_CEDE hypercall) threads 2 and 3 in the
> idle loop and the hypervisor will automatically switch to SMT2 for that
> core (independent of other cores). The opposite is not true, so if
> threads 0 & 1 are idle and 2 & 3 are active, we will stay in SMT4 mode.
>
> Similarly if thread 0 is active and threads 1, 2 & 3 are idle, we'll go
> into SMT1 mode.
>
> If we can get the core into a lower SMT mode (SMT1 is best), the threads
> will perform better (since they share less core resources). Hence when
> we have idle threads, we want them to be the higher ones.
Just out of curiosity, is this a hardware constraint or a hypervisor
constraint?
> So to answer your question, threads 2 and 3 aren't weaker than the other
> threads when in SMT4 mode. It's that if we idle threads 2 & 3, threads
> 0 & 1 will speed up since we'll move to SMT2 mode.
>
> I'm pretty vague on linux scheduler details, so I'm a bit at sea as to
> how to solve this. Can you suggest any mechanisms we currently have in
> the kernel to reflect these properties, or do you think we need to
> develop something new? If so, any pointers as to where we should look?
Well there currently isn't one, and I've been telling people to create a
new SD_flag to reflect this and influence the f_b_g() behaviour.
Something like the below perhaps, totally untested and without comments
so that you'll have to reverse engineer and validate my thinking.
There's one fundamental assumption, and one weakness in the
implementation.
---
include/linux/sched.h | 2 +-
kernel/sched_fair.c | 61 +++++++++++++++++++++++++++++++++++++++++++++---
2 files changed, 58 insertions(+), 5 deletions(-)
diff --git a/include/linux/sched.h b/include/linux/sched.h
index 0eef87b..42fa5c6 100644
--- a/include/linux/sched.h
+++ b/include/linux/sched.h
@@ -849,7 +849,7 @@ enum cpu_idle_type {
#define SD_POWERSAVINGS_BALANCE 0x0100 /* Balance for power savings */
#define SD_SHARE_PKG_RESOURCES 0x0200 /* Domain members share cpu pkg resources */
#define SD_SERIALIZE 0x0400 /* Only a single load balancing instance */
-
+#define SD_ASYM_PACKING 0x0800
#define SD_PREFER_SIBLING 0x1000 /* Prefer to place tasks in a sibling domain */
enum powersavings_balance_level {
diff --git a/kernel/sched_fair.c b/kernel/sched_fair.c
index ff7692c..7e42bfe 100644
--- a/kernel/sched_fair.c
+++ b/kernel/sched_fair.c
@@ -2086,6 +2086,7 @@ struct sd_lb_stats {
struct sched_group *this; /* Local group in this sd */
unsigned long total_load; /* Total load of all groups in sd */
unsigned long total_pwr; /* Total power of all groups in sd */
+ unsigned long total_nr_running;
unsigned long avg_load; /* Average load across all groups in sd */
/** Statistics of this group */
@@ -2414,10 +2415,10 @@ static inline void update_sg_lb_stats(struct sched_domain *sd,
int *balance, struct sg_lb_stats *sgs)
{
unsigned long load, max_cpu_load, min_cpu_load;
- int i;
unsigned int balance_cpu = -1, first_idle_cpu = 0;
unsigned long sum_avg_load_per_task;
unsigned long avg_load_per_task;
+ int i;
if (local_group)
balance_cpu = group_first_cpu(group);
@@ -2493,6 +2494,28 @@ static inline void update_sg_lb_stats(struct sched_domain *sd,
DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE);
}
+static int update_sd_pick_busiest(struct sched_domain *sd,
+ struct sd_lb_stats *sds,
+ struct sched_group *sg,
+ struct sg_lb_stats *sgs)
+{
+ if (sgs->sum_nr_running > sgs->group_capacity)
+ return 1;
+
+ if (sgs->group_imb)
+ return 1;
+
+ if ((sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running) {
+ if (!sds->busiest)
+ return 1;
+
+ if (group_first_cpu(sds->busiest) < group_first_cpu(group))
+ return 1;
+ }
+
+ return 0;
+}
+
/**
* update_sd_lb_stats - Update sched_group's statistics for load balancing.
* @sd: sched_domain whose statistics are to be updated.
@@ -2533,6 +2556,7 @@ static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
sds->total_load += sgs.group_load;
sds->total_pwr += group->cpu_power;
+ sds->total_nr_running += sgs.sum_nr_running;
/*
* In case the child domain prefers tasks go to siblings
@@ -2547,9 +2571,8 @@ static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
sds->this = group;
sds->this_nr_running = sgs.sum_nr_running;
sds->this_load_per_task = sgs.sum_weighted_load;
- } else if (sgs.avg_load > sds->max_load &&
- (sgs.sum_nr_running > sgs.group_capacity ||
- sgs.group_imb)) {
+ } else if (sgs.avg_load >= sds->max_load &&
+ update_sd_pick_busiest(sd, sds, group, &sgs)) {
sds->max_load = sgs.avg_load;
sds->busiest = group;
sds->busiest_nr_running = sgs.sum_nr_running;
@@ -2562,6 +2585,33 @@ static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
} while (group != sd->groups);
}
+static int check_asym_packing(struct sched_domain *sd,
+ struct sd_lb_stats *sds,
+ int cpu, unsigned long *imbalance)
+{
+ int i, cpu, busiest_cpu;
+
+ if (!(sd->flags & SD_ASYM_PACKING))
+ return 0;
+
+ if (!sds->busiest)
+ return 0;
+
+ i = 0;
+ busiest_cpu = group_first_cpu(sds->busiest);
+ for_each_cpu(cpu, sched_domain_span(sd)) {
+ i++;
+ if (cpu == busiest_cpu)
+ break;
+ }
+
+ if (sds->total_nr_running > i)
+ return 0;
+
+ *imbalance = sds->max_load;
+ return 1;
+}
+
/**
* fix_small_imbalance - Calculate the minor imbalance that exists
* amongst the groups of a sched_domain, during
@@ -2761,6 +2811,9 @@ find_busiest_group(struct sched_domain *sd, int this_cpu,
return sds.busiest;
out_balanced:
+ if (check_asym_packing(sd, &sds, this_cpu, imbalance))
+ return sds.busiest;
+
/*
* There is no obvious imbalance. But check if we can do some balancing
* to save power.
On Thu, 2010-02-18 at 14:17 +0100, Peter Zijlstra wrote:
>
> There's one fundamental assumption, and one weakness in the
> implementation.
>
Aside from bugs and the like.. ;-)
Sorry for the slow reply, was on vacation. Mikey seems to have answered
pretty well though.
>>> That is, unless these threads 2 and 3 really are _that_ weak, at which
>>> point one wonders why IBM bothered with the silicon ;-)
>>>
>> Peter,
>>
>> 2 & 3 aren't weaker than 0 & 1 but....
>>
>> The core has dynamic SMT mode switching which is controlled by the
>> hypervisor (IBM's PHYP). There are 3 SMT modes:
>> SMT1 uses thread 0
>> SMT2 uses threads 0 & 1
>> SMT4 uses threads 0, 1, 2 & 3
>> When in any particular SMT mode, all threads have the same performance
>> as each other (ie. at any moment in time, all threads perform the same).
>>
>> The SMT mode switching works such that when linux has threads 2 & 3 idle
>> and 0 & 1 active, it will cede (H_CEDE hypercall) threads 2 and 3 in the
>> idle loop and the hypervisor will automatically switch to SMT2 for that
>> core (independent of other cores). The opposite is not true, so if
>> threads 0 & 1 are idle and 2 & 3 are active, we will stay in SMT4 mode.
>>
>> Similarly if thread 0 is active and threads 1, 2 & 3 are idle, we'll go
>> into SMT1 mode.
>>
>> If we can get the core into a lower SMT mode (SMT1 is best), the threads
>> will perform better (since they share less core resources). Hence when
>> we have idle threads, we want them to be the higher ones.
>>
>
> Just out of curiosity, is this a hardware constraint or a hypervisor
> constraint?
>
hardware
>
>> So to answer your question, threads 2 and 3 aren't weaker than the other
>> threads when in SMT4 mode. It's that if we idle threads 2 & 3, threads
>> 0 & 1 will speed up since we'll move to SMT2 mode.
>>
>> I'm pretty vague on linux scheduler details, so I'm a bit at sea as to
>> how to solve this. Can you suggest any mechanisms we currently have in
>> the kernel to reflect these properties, or do you think we need to
>> develop something new? If so, any pointers as to where we should look?
>>
>
>
Since the threads speed up we'd need to change their weights at runtime
regardless of placement. It just seems to make sense to let the changed
weights affect placement naturally at the same time.
> Well there currently isn't one, and I've been telling people to create a
> new SD_flag to reflect this and influence the f_b_g() behaviour.
>
> Something like the below perhaps, totally untested and without comments
> so that you'll have to reverse engineer and validate my thinking.
>
> There's one fundamental assumption, and one weakness in the
> implementation.
>
I'm going to guess the weakness is that it doesn't adjust the cpu power
so tasks running in SMT1 mode actually get more than they account for?
What's the assumption?
On Thu, 2010-02-18 at 10:28 -0600, Joel Schopp wrote:
> > There's one fundamental assumption, and one weakness in the
> > implementation.
> >
> I'm going to guess the weakness is that it doesn't adjust the cpu power
> so tasks running in SMT1 mode actually get more than they account for?
No, but you're right, if these SMTx modes are running at different
frequencies then yes that needs to happen as well.
The weakness is failing to do the right thing in the presence of a
'strategically' placed RT task.
Suppose:
Sibling0, Sibling1, Sibling2, Sibling3
idle OTHER OTHER FIFO
it might not manage to migrate a task to 0 because it ends up selecting
3 as busiest. It doesn't at all influence RT placement, but it does look
at nr_running (which does include RT tasks)
> What's the assumption?
That cpu_of(Sibling n) < cpu_of(Sibling n+1)
> On Thu, 2010-02-18 at 09:20 +1100, Michael Neuling wrote:
> > > Suppose for a moment we have 2 threads (hot-unplugged thread 1 and 3, we
> > > can construct an equivalent but more complex example for 4 threads), and
> > > we have 4 tasks, 3 SCHED_OTHER of equal nice level and 1 SCHED_FIFO, the
> > > SCHED_FIFO task will consume exactly 50% walltime of whatever cpu it
> > > ends up on.
> > >
> > > In that situation, provided that each cpu's cpu_power is of equal
> > > measure, scale_rt_power() ensures that we run 2 SCHED_OTHER tasks on the
> > > cpu that doesn't run the RT task, and 1 SCHED_OTHER task next to the RT
> > > task, so that each task consumes 50%, which is all fair and proper.
> > >
> > > However, if you do the above, thread 0 will have +75% = 1.75 and thread
> > > 2 will have -75% = 0.25, then if the RT task will land on thread 0,
> > > we'll be having: 0.875 vs 0.25, or on thread 3, 1.75 vs 0.125. In either
> > > case thread 0 will receive too many (if not all) SCHED_OTHER tasks.
> > >
> > > That is, unless these threads 2 and 3 really are _that_ weak, at which
> > > point one wonders why IBM bothered with the silicon ;-)
> >
> > Peter,
> >
> > 2 & 3 aren't weaker than 0 & 1 but....
> >
> > The core has dynamic SMT mode switching which is controlled by the
> > hypervisor (IBM's PHYP). There are 3 SMT modes:
> > SMT1 uses thread 0
> > SMT2 uses threads 0 & 1
> > SMT4 uses threads 0, 1, 2 & 3
> > When in any particular SMT mode, all threads have the same performance
> > as each other (ie. at any moment in time, all threads perform the same).
> >
> > The SMT mode switching works such that when linux has threads 2 & 3 idle
> > and 0 & 1 active, it will cede (H_CEDE hypercall) threads 2 and 3 in the
> > idle loop and the hypervisor will automatically switch to SMT2 for that
> > core (independent of other cores). The opposite is not true, so if
> > threads 0 & 1 are idle and 2 & 3 are active, we will stay in SMT4 mode.
> >
> > Similarly if thread 0 is active and threads 1, 2 & 3 are idle, we'll go
> > into SMT1 mode.
> >
> > If we can get the core into a lower SMT mode (SMT1 is best), the threads
> > will perform better (since they share less core resources). Hence when
> > we have idle threads, we want them to be the higher ones.
>
> Just out of curiosity, is this a hardware constraint or a hypervisor
> constraint?
>
> > So to answer your question, threads 2 and 3 aren't weaker than the other
> > threads when in SMT4 mode. It's that if we idle threads 2 & 3, threads
> > 0 & 1 will speed up since we'll move to SMT2 mode.
> >
> > I'm pretty vague on linux scheduler details, so I'm a bit at sea as to
> > how to solve this. Can you suggest any mechanisms we currently have in
> > the kernel to reflect these properties, or do you think we need to
> > develop something new? If so, any pointers as to where we should look?
>
> Well there currently isn't one, and I've been telling people to create a
> new SD_flag to reflect this and influence the f_b_g() behaviour.
>
> Something like the below perhaps, totally untested and without comments
> so that you'll have to reverse engineer and validate my thinking.
>
> There's one fundamental assumption, and one weakness in the
> implementation.
Thanks for the help.
I'm still trying to get up to speed with how this works but while trying
to cleanup and compile your patch, I had some simple questions below...
>
> ---
>
> include/linux/sched.h | 2 +-
> kernel/sched_fair.c | 61 +++++++++++++++++++++++++++++++++++++++++++++--
-
> 2 files changed, 58 insertions(+), 5 deletions(-)
>
> diff --git a/include/linux/sched.h b/include/linux/sched.h
> index 0eef87b..42fa5c6 100644
> --- a/include/linux/sched.h
> +++ b/include/linux/sched.h
> @@ -849,7 +849,7 @@ enum cpu_idle_type {
> #define SD_POWERSAVINGS_BALANCE 0x0100 /* Balance for power savings */
> #define SD_SHARE_PKG_RESOURCES 0x0200 /* Domain members share cpu pkg
resources */
> #define SD_SERIALIZE 0x0400 /* Only a single load balancing instanc
e */
> -
> +#define SD_ASYM_PACKING 0x0800
Would we eventually add this to SD_SIBLING_INIT in a arch specific hook,
or is this ok to add it generically?
> #define SD_PREFER_SIBLING 0x1000 /* Prefer to place tasks in a sibling d
omain */
>
> enum powersavings_balance_level {
> diff --git a/kernel/sched_fair.c b/kernel/sched_fair.c
> index ff7692c..7e42bfe 100644
> --- a/kernel/sched_fair.c
> +++ b/kernel/sched_fair.c
> @@ -2086,6 +2086,7 @@ struct sd_lb_stats {
> struct sched_group *this; /* Local group in this sd */
> unsigned long total_load; /* Total load of all groups in sd */
> unsigned long total_pwr; /* Total power of all groups in sd */
> + unsigned long total_nr_running;
> unsigned long avg_load; /* Average load across all groups in sd */
>
> /** Statistics of this group */
> @@ -2414,10 +2415,10 @@ static inline void update_sg_lb_stats(struct sched_do
main *sd,
> int *balance, struct sg_lb_stats *sgs)
> {
> unsigned long load, max_cpu_load, min_cpu_load;
> - int i;
> unsigned int balance_cpu = -1, first_idle_cpu = 0;
> unsigned long sum_avg_load_per_task;
> unsigned long avg_load_per_task;
> + int i;
>
> if (local_group)
> balance_cpu = group_first_cpu(group);
> @@ -2493,6 +2494,28 @@ static inline void update_sg_lb_stats(struct sched_dom
ain *sd,
> DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE);
> }
>
> +static int update_sd_pick_busiest(struct sched_domain *sd,
> + struct sd_lb_stats *sds,
> + struct sched_group *sg,
> + struct sg_lb_stats *sgs)
> +{
> + if (sgs->sum_nr_running > sgs->group_capacity)
> + return 1;
> +
> + if (sgs->group_imb)
> + return 1;
> +
> + if ((sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running) {
> + if (!sds->busiest)
> + return 1;
> +
> + if (group_first_cpu(sds->busiest) < group_first_cpu(group))
"group" => "sg" here? (I get a compile error otherwise)
> + return 1;
> + }
> +
> + return 0;
> +}
> +
> /**
> * update_sd_lb_stats - Update sched_group's statistics for load balancing.
> * @sd: sched_domain whose statistics are to be updated.
> @@ -2533,6 +2556,7 @@ static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
>
> sds->total_load += sgs.group_load;
> sds->total_pwr += group->cpu_power;
> + sds->total_nr_running += sgs.sum_nr_running;
>
> /*
> * In case the child domain prefers tasks go to siblings
> @@ -2547,9 +2571,8 @@ static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
> sds->this = group;
> sds->this_nr_running = sgs.sum_nr_running;
> sds->this_load_per_task = sgs.sum_weighted_load;
> - } else if (sgs.avg_load > sds->max_load &&
> - (sgs.sum_nr_running > sgs.group_capacity ||
> - sgs.group_imb)) {
> + } else if (sgs.avg_load >= sds->max_load &&
> + update_sd_pick_busiest(sd, sds, group, &sgs)) {
> sds->max_load = sgs.avg_load;
> sds->busiest = group;
> sds->busiest_nr_running = sgs.sum_nr_running;
> @@ -2562,6 +2585,33 @@ static inline void update_sd_lb_stats(struct sched_dom
ain *sd, int this_cpu,
> } while (group != sd->groups);
> }
>
> +static int check_asym_packing(struct sched_domain *sd,
> + struct sd_lb_stats *sds,
> + int cpu, unsigned long *imbalance)
> +{
> + int i, cpu, busiest_cpu;
Redefining cpu here. Looks like the cpu parameter is not really needed?
> +
> + if (!(sd->flags & SD_ASYM_PACKING))
> + return 0;
> +
> + if (!sds->busiest)
> + return 0;
> +
> + i = 0;
> + busiest_cpu = group_first_cpu(sds->busiest);
> + for_each_cpu(cpu, sched_domain_span(sd)) {
> + i++;
> + if (cpu == busiest_cpu)
> + break;
> + }
> +
> + if (sds->total_nr_running > i)
> + return 0;
> +
> + *imbalance = sds->max_load;
> + return 1;
> +}
> +
> /**
> * fix_small_imbalance - Calculate the minor imbalance that exists
> * amongst the groups of a sched_domain, during
> @@ -2761,6 +2811,9 @@ find_busiest_group(struct sched_domain *sd, int this_cp
u,
> return sds.busiest;
>
> out_balanced:
> + if (check_asym_packing(sd, &sds, this_cpu, imbalance))
> + return sds.busiest;
> +
> /*
> * There is no obvious imbalance. But check if we can do some balancing
> * to save power.
>
>
On Fri, 2010-02-19 at 17:05 +1100, Michael Neuling wrote:
> > include/linux/sched.h | 2 +-
> > kernel/sched_fair.c | 61 +++++++++++++++++++++++++++++++++++++++++++++--
> -
> > 2 files changed, 58 insertions(+), 5 deletions(-)
> >
> > diff --git a/include/linux/sched.h b/include/linux/sched.h
> > index 0eef87b..42fa5c6 100644
> > --- a/include/linux/sched.h
> > +++ b/include/linux/sched.h
> > @@ -849,7 +849,7 @@ enum cpu_idle_type {
> > #define SD_POWERSAVINGS_BALANCE 0x0100 /* Balance for power savings */
> > #define SD_SHARE_PKG_RESOURCES 0x0200 /* Domain members share cpu pkg
> resources */
> > #define SD_SERIALIZE 0x0400 /* Only a single load balancing instanc
> e */
> > -
> > +#define SD_ASYM_PACKING 0x0800
>
> Would we eventually add this to SD_SIBLING_INIT in a arch specific hook,
> or is this ok to add it generically?
I'd think we'd want to keep that limited to architectures that actually
need it.
>
> > +static int update_sd_pick_busiest(struct sched_domain *sd,
> > + struct sd_lb_stats *sds,
> > + struct sched_group *sg,
> > + struct sg_lb_stats *sgs)
> > +{
> > + if (sgs->sum_nr_running > sgs->group_capacity)
> > + return 1;
> > +
> > + if (sgs->group_imb)
> > + return 1;
> > +
> > + if ((sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running) {
> > + if (!sds->busiest)
> > + return 1;
> > +
> > + if (group_first_cpu(sds->busiest) < group_first_cpu(group))
>
> "group" => "sg" here? (I get a compile error otherwise)
Oh, quite ;-)
> > +static int check_asym_packing(struct sched_domain *sd,
> > + struct sd_lb_stats *sds,
> > + int cpu, unsigned long *imbalance)
> > +{
> > + int i, cpu, busiest_cpu;
>
> Redefining cpu here. Looks like the cpu parameter is not really needed?
Seems that way indeed, I went back and forth a few times on the actual
implementation of this function (which started out live as a copy of
check_power_save_busiest_group), its amazing there were only these two
compile glitches ;-)
> > +
> > + if (!(sd->flags & SD_ASYM_PACKING))
> > + return 0;
> > +
> > + if (!sds->busiest)
> > + return 0;
> > +
> > + i = 0;
> > + busiest_cpu = group_first_cpu(sds->busiest);
> > + for_each_cpu(cpu, sched_domain_span(sd)) {
> > + i++;
> > + if (cpu == busiest_cpu)
> > + break;
> > + }
> > +
> > + if (sds->total_nr_running > i)
> > + return 0;
> > +
> > + *imbalance = sds->max_load;
> > + return 1;
> > +}
In message <1266573672.1806.70.camel@laptop> you wrote:
> On Fri, 2010-02-19 at 17:05 +1100, Michael Neuling wrote:
> > > include/linux/sched.h | 2 +-
> > > kernel/sched_fair.c | 61 +++++++++++++++++++++++++++++++++++++++++++
++--
> > -
> > > 2 files changed, 58 insertions(+), 5 deletions(-)
> > >
> > > diff --git a/include/linux/sched.h b/include/linux/sched.h
> > > index 0eef87b..42fa5c6 100644
> > > --- a/include/linux/sched.h
> > > +++ b/include/linux/sched.h
> > > @@ -849,7 +849,7 @@ enum cpu_idle_type {
> > > #define SD_POWERSAVINGS_BALANCE 0x0100 /* Balance for power savings */
> > > #define SD_SHARE_PKG_RESOURCES 0x0200 /* Domain members share cpu pkg
> > resources */
> > > #define SD_SERIALIZE 0x0400 /* Only a single load balancing
instanc
> > e */
> > > -
> > > +#define SD_ASYM_PACKING 0x0800
> >
> > Would we eventually add this to SD_SIBLING_INIT in a arch specific hook,
> > or is this ok to add it generically?
>
> I'd think we'd want to keep that limited to architectures that actually
> need it.
OK
> >
> > > +static int update_sd_pick_busiest(struct sched_domain *sd,
> > > + struct sd_lb_stats *sds,
> > > + struct sched_group *sg,
> > > + struct sg_lb_stats *sgs)
> > > +{
> > > + if (sgs->sum_nr_running > sgs->group_capacity)
> > > + return 1;
> > > +
> > > + if (sgs->group_imb)
> > > + return 1;
> > > +
> > > + if ((sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running) {
> > > + if (!sds->busiest)
> > > + return 1;
> > > +
> > > + if (group_first_cpu(sds->busiest) < group_first_cpu(group))
> >
> > "group" => "sg" here? (I get a compile error otherwise)
>
> Oh, quite ;-)
>
> > > +static int check_asym_packing(struct sched_domain *sd,
> > > + struct sd_lb_stats *sds,
> > > + int cpu, unsigned long *imbalance)
> > > +{
> > > + int i, cpu, busiest_cpu;
> >
> > Redefining cpu here. Looks like the cpu parameter is not really needed?
>
> Seems that way indeed, I went back and forth a few times on the actual
> implementation of this function (which started out live as a copy of
> check_power_save_busiest_group), its amazing there were only these two
> compile glitches ;-)
:-)
Below are the cleanups + the arch specific bits. It doesn't change your
logic at all. Obviously the PPC arch bits would need to be split into a
separate patch.
Compiles and boots against linux-next.
Mikey
---
arch/powerpc/include/asm/cputable.h | 3 +
arch/powerpc/kernel/process.c | 7 +++
include/linux/sched.h | 4 +-
include/linux/topology.h | 1
kernel/sched_fair.c | 64 ++++++++++++++++++++++++++++++++++--
5 files changed, 74 insertions(+), 5 deletions(-)
Index: linux-next/arch/powerpc/include/asm/cputable.h
===================================================================
--- linux-next.orig/arch/powerpc/include/asm/cputable.h
+++ linux-next/arch/powerpc/include/asm/cputable.h
@@ -195,6 +195,7 @@ extern const char *powerpc_base_platform
#define CPU_FTR_SAO LONG_ASM_CONST(0x0020000000000000)
#define CPU_FTR_CP_USE_DCBTZ LONG_ASM_CONST(0x0040000000000000)
#define CPU_FTR_UNALIGNED_LD_STD LONG_ASM_CONST(0x0080000000000000)
+#define CPU_FTR_ASYM_SMT4 LONG_ASM_CONST(0x0100000000000000)
#ifndef __ASSEMBLY__
@@ -409,7 +410,7 @@ extern const char *powerpc_base_platform
CPU_FTR_MMCRA | CPU_FTR_SMT | \
CPU_FTR_COHERENT_ICACHE | CPU_FTR_LOCKLESS_TLBIE | \
CPU_FTR_PURR | CPU_FTR_SPURR | CPU_FTR_REAL_LE | \
- CPU_FTR_DSCR | CPU_FTR_SAO)
+ CPU_FTR_DSCR | CPU_FTR_SAO | CPU_FTR_ASYM_SMT4)
#define CPU_FTRS_CELL (CPU_FTR_USE_TB | CPU_FTR_LWSYNC | \
CPU_FTR_PPCAS_ARCH_V2 | CPU_FTR_CTRL | \
CPU_FTR_ALTIVEC_COMP | CPU_FTR_MMCRA | CPU_FTR_SMT | \
Index: linux-next/arch/powerpc/kernel/process.c
===================================================================
--- linux-next.orig/arch/powerpc/kernel/process.c
+++ linux-next/arch/powerpc/kernel/process.c
@@ -1265,3 +1265,10 @@ unsigned long randomize_et_dyn(unsigned
return ret;
}
+
+int arch_sd_asym_packing(void)
+{
+ if (cpu_has_feature(CPU_FTR_ASYM_SMT4))
+ return SD_ASYM_PACKING;
+ return 0;
+}
Index: linux-next/include/linux/sched.h
===================================================================
--- linux-next.orig/include/linux/sched.h
+++ linux-next/include/linux/sched.h
@@ -849,7 +849,7 @@ enum cpu_idle_type {
#define SD_POWERSAVINGS_BALANCE 0x0100 /* Balance for power savings */
#define SD_SHARE_PKG_RESOURCES 0x0200 /* Domain members share cpu pkg resources */
#define SD_SERIALIZE 0x0400 /* Only a single load balancing instance */
-
+#define SD_ASYM_PACKING 0x0800 /* Asymetric SMT packing */
#define SD_PREFER_SIBLING 0x1000 /* Prefer to place tasks in a sibling domain */
enum powersavings_balance_level {
@@ -881,6 +881,8 @@ static inline int sd_balance_for_package
return SD_PREFER_SIBLING;
}
+extern int arch_sd_asym_packing(void);
+
/*
* Optimise SD flags for power savings:
* SD_BALANCE_NEWIDLE helps agressive task consolidation and power savings.
Index: linux-next/include/linux/topology.h
===================================================================
--- linux-next.orig/include/linux/topology.h
+++ linux-next/include/linux/topology.h
@@ -102,6 +102,7 @@ int arch_update_cpu_topology(void);
| 1*SD_SHARE_PKG_RESOURCES \
| 0*SD_SERIALIZE \
| 0*SD_PREFER_SIBLING \
+ | arch_sd_asym_packing() \
, \
.last_balance = jiffies, \
.balance_interval = 1, \
Index: linux-next/kernel/sched_fair.c
===================================================================
--- linux-next.orig/kernel/sched_fair.c
+++ linux-next/kernel/sched_fair.c
@@ -2086,6 +2086,7 @@ struct sd_lb_stats {
struct sched_group *this; /* Local group in this sd */
unsigned long total_load; /* Total load of all groups in sd */
unsigned long total_pwr; /* Total power of all groups in sd */
+ unsigned long total_nr_running;
unsigned long avg_load; /* Average load across all groups in sd */
/** Statistics of this group */
@@ -2493,6 +2494,28 @@ static inline void update_sg_lb_stats(st
DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE);
}
+static int update_sd_pick_busiest(struct sched_domain *sd,
+ struct sd_lb_stats *sds,
+ struct sched_group *sg,
+ struct sg_lb_stats *sgs)
+{
+ if (sgs->sum_nr_running > sgs->group_capacity)
+ return 1;
+
+ if (sgs->group_imb)
+ return 1;
+
+ if ((sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running) {
+ if (!sds->busiest)
+ return 1;
+
+ if (group_first_cpu(sds->busiest) < group_first_cpu(sg))
+ return 1;
+ }
+
+ return 0;
+}
+
/**
* update_sd_lb_stats - Update sched_group's statistics for load balancing.
* @sd: sched_domain whose statistics are to be updated.
@@ -2533,6 +2556,7 @@ static inline void update_sd_lb_stats(st
sds->total_load += sgs.group_load;
sds->total_pwr += group->cpu_power;
+ sds->total_nr_running += sgs.sum_nr_running;
/*
* In case the child domain prefers tasks go to siblings
@@ -2547,9 +2571,8 @@ static inline void update_sd_lb_stats(st
sds->this = group;
sds->this_nr_running = sgs.sum_nr_running;
sds->this_load_per_task = sgs.sum_weighted_load;
- } else if (sgs.avg_load > sds->max_load &&
- (sgs.sum_nr_running > sgs.group_capacity ||
- sgs.group_imb)) {
+ } else if (sgs.avg_load >= sds->max_load &&
+ update_sd_pick_busiest(sd, sds, group, &sgs)) {
sds->max_load = sgs.avg_load;
sds->busiest = group;
sds->busiest_nr_running = sgs.sum_nr_running;
@@ -2562,6 +2585,38 @@ static inline void update_sd_lb_stats(st
} while (group != sd->groups);
}
+int __weak sd_asym_packing_arch(void)
+{
+ return 0;
+}
+
+static int check_asym_packing(struct sched_domain *sd,
+ struct sd_lb_stats *sds,
+ unsigned long *imbalance)
+{
+ int i, cpu, busiest_cpu;
+
+ if (!(sd->flags & SD_ASYM_PACKING))
+ return 0;
+
+ if (!sds->busiest)
+ return 0;
+
+ i = 0;
+ busiest_cpu = group_first_cpu(sds->busiest);
+ for_each_cpu(cpu, sched_domain_span(sd)) {
+ i++;
+ if (cpu == busiest_cpu)
+ break;
+ }
+
+ if (sds->total_nr_running > i)
+ return 0;
+
+ *imbalance = sds->max_load;
+ return 1;
+}
+
/**
* fix_small_imbalance - Calculate the minor imbalance that exists
* amongst the groups of a sched_domain, during
@@ -2761,6 +2816,9 @@ find_busiest_group(struct sched_domain *
return sds.busiest;
out_balanced:
+ if (check_asym_packing(sd, &sds, imbalance))
+ return sds.busiest;
+
/*
* There is no obvious imbalance. But check if we can do some balancing
* to save power.
In message <[email protected]> you wrote:
> In message <1266573672.1806.70.camel@laptop> you wrote:
> > On Fri, 2010-02-19 at 17:05 +1100, Michael Neuling wrote:
> > > > include/linux/sched.h | 2 +-
> > > > kernel/sched_fair.c | 61 +++++++++++++++++++++++++++++++++++++++++
++
> ++--
> > > -
> > > > 2 files changed, 58 insertions(+), 5 deletions(-)
> > > >
> > > > diff --git a/include/linux/sched.h b/include/linux/sched.h
> > > > index 0eef87b..42fa5c6 100644
> > > > --- a/include/linux/sched.h
> > > > +++ b/include/linux/sched.h
> > > > @@ -849,7 +849,7 @@ enum cpu_idle_type {
> > > > #define SD_POWERSAVINGS_BALANCE 0x0100 /* Balance for power sa
vings */
> > > > #define SD_SHARE_PKG_RESOURCES 0x0200 /* Domain members share cpu pkg
> > > resources */
> > > > #define SD_SERIALIZE 0x0400 /* Only a single load balancing
> instanc
> > > e */
> > > > -
> > > > +#define SD_ASYM_PACKING 0x0800
> > >
> > > Would we eventually add this to SD_SIBLING_INIT in a arch specific hook,
> > > or is this ok to add it generically?
> >
> > I'd think we'd want to keep that limited to architectures that actually
> > need it.
>
> OK
>
> > >
> > > > +static int update_sd_pick_busiest(struct sched_domain *sd,
> > > > + struct sd_lb_stats *sds,
> > > > + struct sched_group *sg,
> > > > + struct sg_lb_stats *sgs)
> > > > +{
> > > > + if (sgs->sum_nr_running > sgs->group_capacity)
> > > > + return 1;
> > > > +
> > > > + if (sgs->group_imb)
> > > > + return 1;
> > > > +
> > > > + if ((sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running) {
> > > > + if (!sds->busiest)
> > > > + return 1;
> > > > +
> > > > + if (group_first_cpu(sds->busiest) < group_first_cpu(gro
up))
> > >
> > > "group" => "sg" here? (I get a compile error otherwise)
> >
> > Oh, quite ;-)
> >
> > > > +static int check_asym_packing(struct sched_domain *sd,
> > > > + struct sd_lb_stats *sds,
> > > > + int cpu, unsigned long *imbalance)
> > > > +{
> > > > + int i, cpu, busiest_cpu;
> > >
> > > Redefining cpu here. Looks like the cpu parameter is not really needed?
> >
> > Seems that way indeed, I went back and forth a few times on the actual
> > implementation of this function (which started out live as a copy of
> > check_power_save_busiest_group), its amazing there were only these two
> > compile glitches ;-)
>
> :-)
>
> Below are the cleanups + the arch specific bits. It doesn't change your
> logic at all. Obviously the PPC arch bits would need to be split into a
> separate patch.
>
> Compiles and boots against linux-next.
I have some comments on the code inline but...
So when I run this, I don't get processes pulled down to the lower
threads. A simple test case of running 1 CPU intensive process at
SCHED_OTHER on a machine with 2 way SMT system (a POWER6 but enabling
SD_ASYM_PACKING). The single processes doesn't move to lower threads as
I'd hope.
Also, are you sure you want to put this in generic code? It seem to be
quite POWER7 specific functionality, so would be logically better in
arch/powerpc. I guess some other arch *might* need it, but seems
unlikely.
> Mikey
> ---
> arch/powerpc/include/asm/cputable.h | 3 +
> arch/powerpc/kernel/process.c | 7 +++
> include/linux/sched.h | 4 +-
> include/linux/topology.h | 1
> kernel/sched_fair.c | 64 +++++++++++++++++++++++++++++++++
+--
> 5 files changed, 74 insertions(+), 5 deletions(-)
>
> Index: linux-next/arch/powerpc/include/asm/cputable.h
> ===================================================================
> --- linux-next.orig/arch/powerpc/include/asm/cputable.h
> +++ linux-next/arch/powerpc/include/asm/cputable.h
> @@ -195,6 +195,7 @@ extern const char *powerpc_base_platform
> #define CPU_FTR_SAO LONG_ASM_CONST(0x0020000000000000)
> #define CPU_FTR_CP_USE_DCBTZ LONG_ASM_CONST(0x0040000000000000)
> #define CPU_FTR_UNALIGNED_LD_STD LONG_ASM_CONST(0x0080000000000000)
> +#define CPU_FTR_ASYM_SMT4 LONG_ASM_CONST(0x0100000000000000)
>
> #ifndef __ASSEMBLY__
>
> @@ -409,7 +410,7 @@ extern const char *powerpc_base_platform
> CPU_FTR_MMCRA | CPU_FTR_SMT | \
> CPU_FTR_COHERENT_ICACHE | CPU_FTR_LOCKLESS_TLBIE | \
> CPU_FTR_PURR | CPU_FTR_SPURR | CPU_FTR_REAL_LE | \
> - CPU_FTR_DSCR | CPU_FTR_SAO)
> + CPU_FTR_DSCR | CPU_FTR_SAO | CPU_FTR_ASYM_SMT4)
> #define CPU_FTRS_CELL (CPU_FTR_USE_TB | CPU_FTR_LWSYNC | \
> CPU_FTR_PPCAS_ARCH_V2 | CPU_FTR_CTRL | \
> CPU_FTR_ALTIVEC_COMP | CPU_FTR_MMCRA | CPU_FTR_SMT | \
> Index: linux-next/arch/powerpc/kernel/process.c
> ===================================================================
> --- linux-next.orig/arch/powerpc/kernel/process.c
> +++ linux-next/arch/powerpc/kernel/process.c
> @@ -1265,3 +1265,10 @@ unsigned long randomize_et_dyn(unsigned
>
> return ret;
> }
> +
> +int arch_sd_asym_packing(void)
> +{
> + if (cpu_has_feature(CPU_FTR_ASYM_SMT4))
> + return SD_ASYM_PACKING;
> + return 0;
> +}
> Index: linux-next/include/linux/sched.h
> ===================================================================
> --- linux-next.orig/include/linux/sched.h
> +++ linux-next/include/linux/sched.h
> @@ -849,7 +849,7 @@ enum cpu_idle_type {
> #define SD_POWERSAVINGS_BALANCE 0x0100 /* Balance for power savings */
> #define SD_SHARE_PKG_RESOURCES 0x0200 /* Domain members share cpu pkg
resources */
> #define SD_SERIALIZE 0x0400 /* Only a single load balancing instanc
e */
> -
> +#define SD_ASYM_PACKING 0x0800 /* Asymetric SMT packing */
> #define SD_PREFER_SIBLING 0x1000 /* Prefer to place tasks in a sibling d
omain */
>
> enum powersavings_balance_level {
> @@ -881,6 +881,8 @@ static inline int sd_balance_for_package
> return SD_PREFER_SIBLING;
> }
>
> +extern int arch_sd_asym_packing(void);
> +
> /*
> * Optimise SD flags for power savings:
> * SD_BALANCE_NEWIDLE helps agressive task consolidation and power savings.
> Index: linux-next/include/linux/topology.h
> ===================================================================
> --- linux-next.orig/include/linux/topology.h
> +++ linux-next/include/linux/topology.h
> @@ -102,6 +102,7 @@ int arch_update_cpu_topology(void);
> | 1*SD_SHARE_PKG_RESOURCES \
> | 0*SD_SERIALIZE \
> | 0*SD_PREFER_SIBLING \
> + | arch_sd_asym_packing() \
> , \
> .last_balance = jiffies, \
> .balance_interval = 1, \
> Index: linux-next/kernel/sched_fair.c
> ===================================================================
> --- linux-next.orig/kernel/sched_fair.c
> +++ linux-next/kernel/sched_fair.c
> @@ -2086,6 +2086,7 @@ struct sd_lb_stats {
> struct sched_group *this; /* Local group in this sd */
> unsigned long total_load; /* Total load of all groups in sd */
> unsigned long total_pwr; /* Total power of all groups in sd */
> + unsigned long total_nr_running;
> unsigned long avg_load; /* Average load across all groups in sd */
>
> /** Statistics of this group */
> @@ -2493,6 +2494,28 @@ static inline void update_sg_lb_stats(st
> DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE);
> }
>
> +static int update_sd_pick_busiest(struct sched_domain *sd,
> + struct sd_lb_stats *sds,
> + struct sched_group *sg,
> + struct sg_lb_stats *sgs)
> +{
> + if (sgs->sum_nr_running > sgs->group_capacity)
> + return 1;
> +
> + if (sgs->group_imb)
> + return 1;
> +
> + if ((sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running) {
If we are asymetric packing...
> + if (!sds->busiest)
> + return 1;
This just seems to be a null pointer check.
>From the tracing I've done, this is always true (always NULL) at this
point so we return here.
> +
> + if (group_first_cpu(sds->busiest) < group_first_cpu(sg))
> + return 1;
I'm a bit lost as to what this is for. Any clues you could provide
would be appreciated. :-)
Is the first cpu in this domain's busiest group before the first cpu in
this group. If, so pick this as the busiest?
Should this be the other way around if we want to pack the busiest to
the first cpu? Mark it as the busiest if it's after (not before).
Is group_first_cpu guaranteed to give us the first physical cpu (ie.
thread 0 in our case) or are these virtualised at this point?
I'm not seeing this hit anyway due to the null pointer check above.
> + }
> +
> + return 0;
> +}
> +
> /**
> * update_sd_lb_stats - Update sched_group's statistics for load balancing.
> * @sd: sched_domain whose statistics are to be updated.
> @@ -2533,6 +2556,7 @@ static inline void update_sd_lb_stats(st
>
> sds->total_load += sgs.group_load;
> sds->total_pwr += group->cpu_power;
> + sds->total_nr_running += sgs.sum_nr_running;
>
> /*
> * In case the child domain prefers tasks go to siblings
> @@ -2547,9 +2571,8 @@ static inline void update_sd_lb_stats(st
> sds->this = group;
> sds->this_nr_running = sgs.sum_nr_running;
> sds->this_load_per_task = sgs.sum_weighted_load;
> - } else if (sgs.avg_load > sds->max_load &&
> - (sgs.sum_nr_running > sgs.group_capacity ||
> - sgs.group_imb)) {
> + } else if (sgs.avg_load >= sds->max_load &&
> + update_sd_pick_busiest(sd, sds, group, &sgs)) {
This is pretty clear. Moving stuff to the new function.
> sds->max_load = sgs.avg_load;
> sds->busiest = group;
> sds->busiest_nr_running = sgs.sum_nr_running;
> @@ -2562,6 +2585,38 @@ static inline void update_sd_lb_stats(st
> } while (group != sd->groups);
> }
>
> +int __weak sd_asym_packing_arch(void)
> +{
> + return 0;
> +}
> +
> +static int check_asym_packing(struct sched_domain *sd,
> + struct sd_lb_stats *sds,
> + unsigned long *imbalance)
> +{
> + int i, cpu, busiest_cpu;
> +
> + if (!(sd->flags & SD_ASYM_PACKING))
> + return 0;
> +
> + if (!sds->busiest)
> + return 0;
> +
> + i = 0;
> + busiest_cpu = group_first_cpu(sds->busiest);
> + for_each_cpu(cpu, sched_domain_span(sd)) {
> + i++;
> + if (cpu == busiest_cpu)
> + break;
> + }
> +
> + if (sds->total_nr_running > i)
> + return 0;
This seems to be the core of the packing logic.
We make sure the busiest_cpu is not past total_nr_running. If it is we
mark as imbalanced. Correct?
It seems if a non zero thread/group had a pile of processes running on
it and a lower thread had much less, this wouldn't fire, but I'm
guessing normal load balancing would kick in that case to fix the
imbalance.
Any corrections to my ramblings appreciated :-)
Thanks again,
Mikey
> +
> + *imbalance = sds->max_load;
> + return 1;
> +}
> +
> /**
> * fix_small_imbalance - Calculate the minor imbalance that exists
> * amongst the groups of a sched_domain, during
> @@ -2761,6 +2816,9 @@ find_busiest_group(struct sched_domain *
> return sds.busiest;
>
> out_balanced:
> + if (check_asym_packing(sd, &sds, imbalance))
> + return sds.busiest;
> +
> /*
> * There is no obvious imbalance. But check if we can do some balancing
> * to save power.
On Tue, 2010-02-23 at 17:08 +1100, Michael Neuling wrote:
> I have some comments on the code inline but...
>
> So when I run this, I don't get processes pulled down to the lower
> threads. A simple test case of running 1 CPU intensive process at
> SCHED_OTHER on a machine with 2 way SMT system (a POWER6 but enabling
> SD_ASYM_PACKING). The single processes doesn't move to lower threads as
> I'd hope.
>
> Also, are you sure you want to put this in generic code? It seem to be
> quite POWER7 specific functionality, so would be logically better in
> arch/powerpc. I guess some other arch *might* need it, but seems
> unlikely.
Well, there are no arch hooks in the load-balancing (aside from the
recent cpu_power stuff, and that really is the wrong thing to poke at
for this), and I did hear some other people express interest in such a
constraint.
Also, load-balancing is complex enough as it is, so I prefer to keep
everything in the generic code where possible, clearly things like
sched_domain creation need arch topology bits, and the arch_scale*
things require other arch information like cpu frequency.
> > @@ -2493,6 +2494,28 @@ static inline void update_sg_lb_stats(st
> > DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE);
> > }
> >
> > +static int update_sd_pick_busiest(struct sched_domain *sd,
> > + struct sd_lb_stats *sds,
> > + struct sched_group *sg,
> > + struct sg_lb_stats *sgs)
> > +{
> > + if (sgs->sum_nr_running > sgs->group_capacity)
> > + return 1;
> > +
> > + if (sgs->group_imb)
> > + return 1;
> > +
> > + if ((sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running) {
>
> If we are asymetric packing...
>
>
> > + if (!sds->busiest)
> > + return 1;
>
> This just seems to be a null pointer check.
>
> From the tracing I've done, this is always true (always NULL) at this
> point so we return here.
Right, so we need to have a busiest group to take a task from, if there
is no busiest yet, take this group.
And in your scenario, with there being only a single task, we'd only hit
this once at most, so yes it makes sense this is always NULL.
> > +
> > + if (group_first_cpu(sds->busiest) < group_first_cpu(sg))
> > + return 1;
>
> I'm a bit lost as to what this is for. Any clues you could provide
> would be appreciated. :-)
>
> Is the first cpu in this domain's busiest group before the first cpu in
> this group. If, so pick this as the busiest?
>
> Should this be the other way around if we want to pack the busiest to
> the first cpu? Mark it as the busiest if it's after (not before).
>
> Is group_first_cpu guaranteed to give us the first physical cpu (ie.
> thread 0 in our case) or are these virtualised at this point?
>
> I'm not seeing this hit anyway due to the null pointer check above.
So this says, if all things being equal, and we already have a busiest,
but this candidate (sg) is higher than the current (busiest) take this
one.
The idea is to move the highest SMT task down.
> > @@ -2562,6 +2585,38 @@ static inline void update_sd_lb_stats(st
> > } while (group != sd->groups);
> > }
> >
> > +int __weak sd_asym_packing_arch(void)
> > +{
> > + return 0;
> > +}
arch_sd_asym_packing() is what you used in topology.h
> > +static int check_asym_packing(struct sched_domain *sd,
> > + struct sd_lb_stats *sds,
> > + unsigned long *imbalance)
> > +{
> > + int i, cpu, busiest_cpu;
> > +
> > + if (!(sd->flags & SD_ASYM_PACKING))
> > + return 0;
> > +
> > + if (!sds->busiest)
> > + return 0;
> > +
> > + i = 0;
> > + busiest_cpu = group_first_cpu(sds->busiest);
> > + for_each_cpu(cpu, sched_domain_span(sd)) {
> > + i++;
> > + if (cpu == busiest_cpu)
> > + break;
> > + }
> > +
> > + if (sds->total_nr_running > i)
> > + return 0;
>
> This seems to be the core of the packing logic.
>
> We make sure the busiest_cpu is not past total_nr_running. If it is we
> mark as imbalanced. Correct?
>
> It seems if a non zero thread/group had a pile of processes running on
> it and a lower thread had much less, this wouldn't fire, but I'm
> guessing normal load balancing would kick in that case to fix the
> imbalance.
>
> Any corrections to my ramblings appreciated :-)
Right, so we're concerned the scenario where there's less tasks than SMT
siblings, if there's more they should all be running and the regular
load-balancer will deal with it.
If there's less the group will normally be balanced and we fall out and
end up in check_asym_packing().
So what I tried doing with that loop is detect if there's a hole in the
packing before busiest. Now that I think about it, what we need to check
is if this_cpu (the removed cpu argument) is idle and less than busiest.
So something like:
static int check_asym_pacing(struct sched_domain *sd,
struct sd_lb_stats *sds,
int this_cpu, unsigned long *imbalance)
{
int busiest_cpu;
if (!(sd->flags & SD_ASYM_PACKING))
return 0;
if (!sds->busiest)
return 0;
busiest_cpu = group_first_cpu(sds->busiest);
if (cpu_rq(this_cpu)->nr_running || this_cpu > busiest_cpu)
return 0;
*imbalance = (sds->max_load * sds->busiest->cpu_power) /
SCHED_LOAD_SCALE;
return 1;
}
Does that make sense?
I still see two problems with this though,.. regular load-balancing only
balances on the first cpu of a domain (see the *balance = 0, condition
in update_sg_lb_stats()), this means that if SMT[12] are idle we'll not
pull properly. Also, nohz balancing might mess this up further.
We could maybe play some games with the balance decision in
update_sg_lb_stats() for SD_ASYM_PACKING domains and idle == CPU_IDLE,
no ideas yet on nohz though.
On Tue, 2010-02-23 at 17:24 +0100, Peter Zijlstra wrote:
>
> busiest_cpu = group_first_cpu(sds->busiest);
> if (cpu_rq(this_cpu)->nr_running || this_cpu > busiest_cpu)
> return 0;
Hmm, we could change the bit in find_busiest_group() to:
if (idle == CPU_IDLE && check_asym_packing())
and skip the nr_running test..
In message <1266942281.11845.521.camel@laptop> you wrote:
> On Tue, 2010-02-23 at 17:08 +1100, Michael Neuling wrote:
>
> > I have some comments on the code inline but...
> >
> > So when I run this, I don't get processes pulled down to the lower
> > threads. A simple test case of running 1 CPU intensive process at
> > SCHED_OTHER on a machine with 2 way SMT system (a POWER6 but enabling
> > SD_ASYM_PACKING). The single processes doesn't move to lower threads as
> > I'd hope.
> >
> > Also, are you sure you want to put this in generic code? It seem to be
> > quite POWER7 specific functionality, so would be logically better in
> > arch/powerpc. I guess some other arch *might* need it, but seems
> > unlikely.
>
> Well, there are no arch hooks in the load-balancing (aside from the
> recent cpu_power stuff, and that really is the wrong thing to poke at
> for this), and I did hear some other people express interest in such a
> constraint.
Interesting. Can I ask, were people interesting in this at the SMT
level or higher in the hierarchy?
> Also, load-balancing is complex enough as it is, so I prefer to keep
> everything in the generic code where possible, clearly things like
> sched_domain creation need arch topology bits, and the arch_scale*
> things require other arch information like cpu frequency.
OK
> > > @@ -2493,6 +2494,28 @@ static inline void update_sg_lb_stats(st
> > > DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE);
> > > }
> > >
> > > +static int update_sd_pick_busiest(struct sched_domain *sd,
> > > + struct sd_lb_stats *sds,
> > > + struct sched_group *sg,
> > > + struct sg_lb_stats *sgs)
> > > +{
> > > + if (sgs->sum_nr_running > sgs->group_capacity)
> > > + return 1;
> > > +
> > > + if (sgs->group_imb)
> > > + return 1;
> > > +
> > > + if ((sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running) {
> >
> > If we are asymetric packing...
> >
> >
> > > + if (!sds->busiest)
> > > + return 1;
> >
> > This just seems to be a null pointer check.
> >
> > From the tracing I've done, this is always true (always NULL) at this
> > point so we return here.
>
> Right, so we need to have a busiest group to take a task from, if there
> is no busiest yet, take this group.
>
> And in your scenario, with there being only a single task, we'd only hit
> this once at most, so yes it makes sense this is always NULL.
OK
>
> > > +
> > > + if (group_first_cpu(sds->busiest) < group_first_cpu(sg))
> > > + return 1;
> >
> > I'm a bit lost as to what this is for. Any clues you could provide
> > would be appreciated. :-)
> >
> > Is the first cpu in this domain's busiest group before the first cpu in
> > this group. If, so pick this as the busiest?
> >
> > Should this be the other way around if we want to pack the busiest to
> > the first cpu? Mark it as the busiest if it's after (not before).
> >
> > Is group_first_cpu guaranteed to give us the first physical cpu (ie.
> > thread 0 in our case) or are these virtualised at this point?
> >
> > I'm not seeing this hit anyway due to the null pointer check above.
>
> So this says, if all things being equal, and we already have a busiest,
> but this candidate (sg) is higher than the current (busiest) take this
> one.
>
> The idea is to move the highest SMT task down.
So in the p7 case, I don't think this is required, as the threads are
all of the same performance when in a given SMT mode. So we don't need
to change the order if the lower groups are busy anyway. It's only once
they became idle that we'd need to rebalanced.
That being said, it probably doesn't hurt?
>
> > > @@ -2562,6 +2585,38 @@ static inline void update_sd_lb_stats(st
> > > } while (group != sd->groups);
> > > }
> > >
> > > +int __weak sd_asym_packing_arch(void)
> > > +{
> > > + return 0;
> > > +}
>
> arch_sd_asym_packing() is what you used in topology.h
Oops, thanks. That would make the function even weaker than I'd
intended :-)
> > > +static int check_asym_packing(struct sched_domain *sd,
> > > + struct sd_lb_stats *sds,
> > > + unsigned long *imbalance)
> > > +{
> > > + int i, cpu, busiest_cpu;
> > > +
> > > + if (!(sd->flags & SD_ASYM_PACKING))
> > > + return 0;
> > > +
> > > + if (!sds->busiest)
> > > + return 0;
> > > +
> > > + i = 0;
> > > + busiest_cpu = group_first_cpu(sds->busiest);
> > > + for_each_cpu(cpu, sched_domain_span(sd)) {
> > > + i++;
> > > + if (cpu == busiest_cpu)
> > > + break;
> > > + }
> > > +
> > > + if (sds->total_nr_running > i)
> > > + return 0;
> >
> > This seems to be the core of the packing logic.
> >
> > We make sure the busiest_cpu is not past total_nr_running. If it is we
> > mark as imbalanced. Correct?
> >
> > It seems if a non zero thread/group had a pile of processes running on
> > it and a lower thread had much less, this wouldn't fire, but I'm
> > guessing normal load balancing would kick in that case to fix the
> > imbalance.
> >
> > Any corrections to my ramblings appreciated :-)
>
> Right, so we're concerned the scenario where there's less tasks than SMT
> siblings, if there's more they should all be running and the regular
> load-balancer will deal with it.
Yes
> If there's less the group will normally be balanced and we fall out and
> end up in check_asym_packing().
>
> So what I tried doing with that loop is detect if there's a hole in the
> packing before busiest. Now that I think about it, what we need to check
> is if this_cpu (the removed cpu argument) is idle and less than busiest.
>
> So something like:
>
> static int check_asym_pacing(struct sched_domain *sd,
> struct sd_lb_stats *sds,
> int this_cpu, unsigned long *imbalance)
> {
> int busiest_cpu;
>
> if (!(sd->flags & SD_ASYM_PACKING))
> return 0;
>
> if (!sds->busiest)
> return 0;
>
> busiest_cpu = group_first_cpu(sds->busiest);
> if (cpu_rq(this_cpu)->nr_running || this_cpu > busiest_cpu)
> return 0;
>
> *imbalance = (sds->max_load * sds->busiest->cpu_power) /
> SCHED_LOAD_SCALE;
> return 1;
> }
>
> Does that make sense?
I think so.
I'm seeing check_asym_packing do the right thing with the simple SMT2
with 1 process case. It marks cpu0 as imbalanced when cpu0 is idle and
cpu1 is busy.
Unfortunately the process doesn't seem to be get migrated down though.
Do we need to give *imbalance a higher value?
FYI this version doesn't use sgs->total_nr_running anymore so I've
removed it.
>
> I still see two problems with this though,.. regular load-balancing only
> balances on the first cpu of a domain (see the *balance = 0, condition
> in update_sg_lb_stats()), this means that if SMT[12] are idle we'll not
> pull properly. Also, nohz balancing might mess this up further.
I do have CONFIG_NO_HZ set but turning it off doesn't help with the
above issue.
> We could maybe play some games with the balance decision in
> update_sg_lb_stats() for SD_ASYM_PACKING domains and idle == CPU_IDLE,
> no ideas yet on nohz though.
OK
<from followup email>
> Hmm, we could change the bit in find_busiest_group() to:
>
> if (idle == CPU_IDLE && check_asym_packing())
>
> and skip the nr_running test..
ok, changed.
This is the patch so far minus the trivial PPC bits.
Thanks!
Mikey
---
include/linux/sched.h | 4 ++
include/linux/topology.h | 1
kernel/sched_fair.c | 65 ++++++++++++++++++++++++++++++++++++++++++++---
3 files changed, 66 insertions(+), 4 deletions(-)
Index: linux-next/include/linux/sched.h
===================================================================
--- linux-next.orig/include/linux/sched.h
+++ linux-next/include/linux/sched.h
@@ -849,7 +849,7 @@ enum cpu_idle_type {
#define SD_POWERSAVINGS_BALANCE 0x0100 /* Balance for power savings */
#define SD_SHARE_PKG_RESOURCES 0x0200 /* Domain members share cpu pkg resources */
#define SD_SERIALIZE 0x0400 /* Only a single load balancing instance */
-
+#define SD_ASYM_PACKING 0x0800 /* Asymmetric SMT packing */
#define SD_PREFER_SIBLING 0x1000 /* Prefer to place tasks in a sibling domain */
enum powersavings_balance_level {
@@ -881,6 +881,8 @@ static inline int sd_balance_for_package
return SD_PREFER_SIBLING;
}
+extern int arch_sd_asym_packing(void);
+
/*
* Optimise SD flags for power savings:
* SD_BALANCE_NEWIDLE helps agressive task consolidation and power savings.
Index: linux-next/include/linux/topology.h
===================================================================
--- linux-next.orig/include/linux/topology.h
+++ linux-next/include/linux/topology.h
@@ -102,6 +102,7 @@ int arch_update_cpu_topology(void);
| 1*SD_SHARE_PKG_RESOURCES \
| 0*SD_SERIALIZE \
| 0*SD_PREFER_SIBLING \
+ | arch_sd_asym_packing() \
, \
.last_balance = jiffies, \
.balance_interval = 1, \
Index: linux-next/kernel/sched_fair.c
===================================================================
--- linux-next.orig/kernel/sched_fair.c
+++ linux-next/kernel/sched_fair.c
@@ -2494,6 +2494,32 @@ static inline void update_sg_lb_stats(st
}
/**
+ * update_sd_pick_busiest - return 1 on busiest group
+ */
+static int update_sd_pick_busiest(struct sched_domain *sd,
+ struct sd_lb_stats *sds,
+ struct sched_group *sg,
+ struct sg_lb_stats *sgs)
+{
+ if (sgs->sum_nr_running > sgs->group_capacity)
+ return 1;
+
+ if (sgs->group_imb)
+ return 1;
+
+ /* Check packing mode for this domain */
+ if ((sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running) {
+ if (!sds->busiest)
+ return 1;
+
+ if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
+ return 1;
+ }
+
+ return 0;
+}
+
+/**
* update_sd_lb_stats - Update sched_group's statistics for load balancing.
* @sd: sched_domain whose statistics are to be updated.
* @this_cpu: Cpu for which load balance is currently performed.
@@ -2547,9 +2573,8 @@ static inline void update_sd_lb_stats(st
sds->this = group;
sds->this_nr_running = sgs.sum_nr_running;
sds->this_load_per_task = sgs.sum_weighted_load;
- } else if (sgs.avg_load > sds->max_load &&
- (sgs.sum_nr_running > sgs.group_capacity ||
- sgs.group_imb)) {
+ } else if (sgs.avg_load >= sds->max_load &&
+ update_sd_pick_busiest(sd, sds, group, &sgs)) {
sds->max_load = sgs.avg_load;
sds->busiest = group;
sds->busiest_nr_running = sgs.sum_nr_running;
@@ -2562,6 +2587,36 @@ static inline void update_sd_lb_stats(st
} while (group != sd->groups);
}
+int __weak arch_sd_asym_packing(void)
+{
+ return 0*SD_ASYM_PACKING;
+}
+
+/**
+ * check_asym_packing - Check to see if we the group is packed into
+ * the sched doman
+ */
+static int check_asym_packing(struct sched_domain *sd,
+ struct sd_lb_stats *sds,
+ int this_cpu, unsigned long *imbalance)
+{
+ int busiest_cpu;
+
+ if (!(sd->flags & SD_ASYM_PACKING))
+ return 0;
+
+ if (!sds->busiest)
+ return 0;
+
+ busiest_cpu = group_first_cpu(sds->busiest);
+ if (this_cpu > busiest_cpu)
+ return 0;
+
+ *imbalance = (sds->max_load * sds->busiest->cpu_power) /
+ SCHED_LOAD_SCALE;
+ return 1;
+}
+
/**
* fix_small_imbalance - Calculate the minor imbalance that exists
* amongst the groups of a sched_domain, during
@@ -2761,6 +2816,10 @@ find_busiest_group(struct sched_domain *
return sds.busiest;
out_balanced:
+ if (idle == CPU_IDLE &&
+ check_asym_packing(sd, &sds, this_cpu, imbalance))
+ return sds.busiest;
+
/*
* There is no obvious imbalance. But check if we can do some balancing
* to save power.
> > If there's less the group will normally be balanced and we fall out and
> > end up in check_asym_packing().
> >
> > So what I tried doing with that loop is detect if there's a hole in the
> > packing before busiest. Now that I think about it, what we need to check
> > is if this_cpu (the removed cpu argument) is idle and less than busiest.
> >
> > So something like:
> >
> > static int check_asym_pacing(struct sched_domain *sd,
> > struct sd_lb_stats *sds,
> > int this_cpu, unsigned long *imbalance)
> > {
> > int busiest_cpu;
> >
> > if (!(sd->flags & SD_ASYM_PACKING))
> > return 0;
> >
> > if (!sds->busiest)
> > return 0;
> >
> > busiest_cpu = group_first_cpu(sds->busiest);
> > if (cpu_rq(this_cpu)->nr_running || this_cpu > busiest_cpu)
> > return 0;
> >
> > *imbalance = (sds->max_load * sds->busiest->cpu_power) /
> > SCHED_LOAD_SCALE;
> > return 1;
> > }
> >
> > Does that make sense?
>
> I think so.
>
> I'm seeing check_asym_packing do the right thing with the simple SMT2
> with 1 process case. It marks cpu0 as imbalanced when cpu0 is idle and
> cpu1 is busy.
>
> Unfortunately the process doesn't seem to be get migrated down though.
> Do we need to give *imbalance a higher value?
So with ego help, I traced this down a bit more.
In my simple test case (SMT2, t0 idle, t1 active) if f_b_g() hits our
new case in check_asym_packing(), load_balance then runs f_b_q().
f_b_q() has this:
if (capacity && rq->nr_running == 1 && wl > imbalance)
continue;
when check_asym_packing() hits, wl = 1783 and imbalance = 1024, so we
continue and busiest remains NULL.
load_balance then does "goto out_balanced" and it doesn't attempt to
move the task.
Based on this and on egos suggestion I pulled in Suresh Siddha patch
from: http://lkml.org/lkml/2010/2/12/352. This fixes the problem. The
process is moved down to t0.
I've only tested SMT2 so far.
Mikey
In message <[email protected]> you wrote:
> > > If there's less the group will normally be balanced and we fall out and
> > > end up in check_asym_packing().
> > >
> > > So what I tried doing with that loop is detect if there's a hole in the
> > > packing before busiest. Now that I think about it, what we need to check
> > > is if this_cpu (the removed cpu argument) is idle and less than busiest.
> > >
> > > So something like:
> > >
> > > static int check_asym_pacing(struct sched_domain *sd,
> > > struct sd_lb_stats *sds,
> > > int this_cpu, unsigned long *imbalance)
> > > {
> > > int busiest_cpu;
> > >
> > > if (!(sd->flags & SD_ASYM_PACKING))
> > > return 0;
> > >
> > > if (!sds->busiest)
> > > return 0;
> > >
> > > busiest_cpu = group_first_cpu(sds->busiest);
> > > if (cpu_rq(this_cpu)->nr_running || this_cpu > busiest_cpu)
> > > return 0;
> > >
> > > *imbalance = (sds->max_load * sds->busiest->cpu_power) /
> > > SCHED_LOAD_SCALE;
> > > return 1;
> > > }
> > >
> > > Does that make sense?
> >
> > I think so.
> >
> > I'm seeing check_asym_packing do the right thing with the simple SMT2
> > with 1 process case. It marks cpu0 as imbalanced when cpu0 is idle and
> > cpu1 is busy.
> >
> > Unfortunately the process doesn't seem to be get migrated down though.
> > Do we need to give *imbalance a higher value?
>
> So with ego help, I traced this down a bit more.
>
> In my simple test case (SMT2, t0 idle, t1 active) if f_b_g() hits our
> new case in check_asym_packing(), load_balance then runs f_b_q().
> f_b_q() has this:
>
> if (capacity && rq->nr_running == 1 && wl > imbalance)
> continue;
>
> when check_asym_packing() hits, wl = 1783 and imbalance = 1024, so we
> continue and busiest remains NULL.
>
> load_balance then does "goto out_balanced" and it doesn't attempt to
> move the task.
>
> Based on this and on egos suggestion I pulled in Suresh Siddha patch
> from: http://lkml.org/lkml/2010/2/12/352. This fixes the problem. The
> process is moved down to t0.
>
> I've only tested SMT2 so far.
SMT4 also works in the simple test case of a single process being pulled
down to thread 0.
As you suspected though, unfortunately this is only working with
CONFIG_NO_HZ off. If I turn NO_HZ on, my single process gets bounced
around the core.
Did you think of any ideas for how to fix the NO_HZ interaction?
Mikey
In message <[email protected]> you wrote:
> > > If there's less the group will normally be balanced and we fall out and
> > > end up in check_asym_packing().
> > >
> > > So what I tried doing with that loop is detect if there's a hole in the
> > > packing before busiest. Now that I think about it, what we need to check
> > > is if this_cpu (the removed cpu argument) is idle and less than busiest.
> > >
> > > So something like:
> > >
> > > static int check_asym_pacing(struct sched_domain *sd,
> > > struct sd_lb_stats *sds,
> > > int this_cpu, unsigned long *imbalance)
> > > {
> > > int busiest_cpu;
> > >
> > > if (!(sd->flags & SD_ASYM_PACKING))
> > > return 0;
> > >
> > > if (!sds->busiest)
> > > return 0;
> > >
> > > busiest_cpu = group_first_cpu(sds->busiest);
> > > if (cpu_rq(this_cpu)->nr_running || this_cpu > busiest_cpu)
> > > return 0;
> > >
> > > *imbalance = (sds->max_load * sds->busiest->cpu_power) /
> > > SCHED_LOAD_SCALE;
> > > return 1;
> > > }
> > >
> > > Does that make sense?
> >
> > I think so.
> >
> > I'm seeing check_asym_packing do the right thing with the simple SMT2
> > with 1 process case. It marks cpu0 as imbalanced when cpu0 is idle and
> > cpu1 is busy.
> >
> > Unfortunately the process doesn't seem to be get migrated down though.
> > Do we need to give *imbalance a higher value?
>
> So with ego help, I traced this down a bit more.
>
> In my simple test case (SMT2, t0 idle, t1 active) if f_b_g() hits our
> new case in check_asym_packing(), load_balance then runs f_b_q().
> f_b_q() has this:
>
> if (capacity && rq->nr_running == 1 && wl > imbalance)
> continue;
>
> when check_asym_packing() hits, wl = 1783 and imbalance = 1024, so we
> continue and busiest remains NULL.
>
> load_balance then does "goto out_balanced" and it doesn't attempt to
> move the task.
>
> Based on this and on egos suggestion I pulled in Suresh Siddha patch
> from: http://lkml.org/lkml/2010/2/12/352. This fixes the problem. The
> process is moved down to t0.
>
> I've only tested SMT2 so far.
I'm finding this SMT2 result to be unreliable. Sometimes it doesn't work
for the simple 1 process case. It seems to change boot to boot.
Sometimes it works as expected with t0 busy and t1 idle, but other times
it's the other way around.
When it doesn't work, check_asym_packing() is still marking processes to
be pulled down but only gets run about 1 in every 4 calls to
load_balance().
For 2 of the other calls to load_balance, idle is CPU_NEWLY_IDLE and
hence check_asym_packing() doesn't get called. This results in
sd->nr_balance_failed being reset. When load_balance is next called and
check_asym_packing() hits, need_active_balance() returns 0 as
sd->nr_balance_failed is too small. This means the migration thread on
t1 is not woken and the process remains there.
So why does thread0 change from NEWLY_IDLE to IDLE and visa versa, when
there is nothing running on it? Is this expected?
This is with next-20100225 which pulled in Ingos tip at
407b4844f2af416cd8c9edd7c44d1545c93a4938 (from 24/2/2010)
Mikey