This is an important point that Mike is raising and it also addresses a
critique that Ingo issued yesterday, namely interactivity and fairness.
The HP scheduler completely separates the per-CPU runqueues and does
not take preemption goodness or alike into account. This can lead to
unfair proportionment of CPU cycles, strong priority inversion and a
potential
lack of interactivity.
Our MQ scheduler does yield the same decision in most cases
(other than defined by some race condition on locks and counter members)
It is not clear that yielding the same decision as the current scheduler
is the ultimate goal to shoot for, but it allows comparision.
Another point to raise is that the current scheduler does a exhaustive
search
for the "best" task to run. It touches every process in the runqueue. this
is
ok if the runqueue length is limited to a very small multiple of the #cpus.
But that is not what high end server systems encounter.
With the rising number of processors, lock contention can quickly become
a bottleneck. If we assume that load (#running-task) increases somewhat
linear with #cpus, the problem gets even worth because not only have I
increased the number of clients but also the lock hold time.
Clinging on to the statement that #running-tasks ~ #cpus, ofcourse saves
you from that dilemma, but not everybody is signing on to this limitation.
MQ and priority-list help in 2 ways.
MQ reduces the average lock holdtime because on average it only inspects
#running-tasks/#cpus tasks to make a local decision. It then goes on to
inspect (#cpus-1) datastructures representing the next best to run tasks
on those remote cpus all without holding the lock, thus substantially
reducing lock contention. Note we still search the entire set of runnable
tasks, however we do it in a distributed collaborative manner.
The only time we deviate from the current scheduler decision is in the
case when two cpus have identified the same remote task as a target for
remote stealing. In that case one will win and the other cpu will continue
looking somewhere else, although there might have been another tasks
on that cpu to steal.
priority list schedulers (PRS) only helps in reducing lock hold time,
which can result in some relieve wrt lock contention, but not a whole lot.
PRS can limit the lists it has to search based on the PROC_CHANGE_PENALTY.
It also keeps 0-counter in a list that is never inspected. One can
even go further and put YIELD tasks in a separate list, given that the
sys_sched_yield already does some optimizations.
The older version (12/00) posted on LSE is functionally equivalent to the
current scheduler.
I will put up another version this week that is based on a bitmask and
which is a bit more agressive in that it ignores the MM goodness boost of 1
which in my books is merely a tie breaker between two task of equal
goodness.
Beyond that we have done some work on cpu pooling, which is to identify
a set of cpus that form a scheduling set. We still consider in
reschedule_idle all cpus for preemption but in schedule it is sufficient
to only schedule within the own set. That again can limit lock hold time
with MQ and we have seen some improvements. We also deploy load balacing.
To summarize, we have taken great care of trying to preserve the current
scheduler semantics and have laid out a path to relax some of the semantics
for further improvements.
I don't believe that the HP scheduler is sufficient since it is lacking
load balacing, which naturally occurs in our MQ scheduler, and it lacks
the interactivity requirements that Ingo pointed out.
Most of these things are discussed in great detail in the writeups under
lse.sourceforge.net/scheduling. I also believe we show there that the
MQ performance for low thread counts is also matching the vanilla case.
I further don't understand the obsession of keeping the scheduler simple.
If there are improvements and I don't believe the MQ is all that
complicated
and its well documented, why not put it in.
Hubertus Franke
Enterprise Linux Group (Mgr), Linux Technology Center (Member Scalability)
email: [email protected]
(w) 914-945-2003 (fax) 914-945-4425 TL: 862-2003
Mike Kravetz <[email protected]> on 04/03/2001 10:47:00 PM
To: Fabio Riccardi <[email protected]>
cc: Mike Kravetz <[email protected]>, Ingo Molnar <[email protected]>,
Hubertus Franke/Watson/IBM@IBMUS, Linux Kernel List
<[email protected]>, Alan Cox <[email protected]>
Subject: Re: a quest for a better scheduler
On Tue, Apr 03, 2001 at 05:18:03PM -0700, Fabio Riccardi wrote:
>
> I have measured the HP and not the "scalability" patch because the two do
more
> or less the same thing and give me the same performance advantages, but
the
> former is a lot simpler and I could port it with no effort on any recent
> kernel.
Actually, there is a significant difference between the HP patch and
the one I developed. In the HP patch, if there is a schedulable task
on the 'local' (current CPU) runqueue it will ignore runnable tasks on
other (remote) runqueues. In the multi-queue patch I developed, the
scheduler always attempts to make the same global scheduling decisions
as the current scheduler.
--
Mike Kravetz [email protected]
IBM Linux Technology Center
On Wed, 4 Apr 2001, Hubertus Franke wrote:
> It is not clear that yielding the same decision as the current
> scheduler is the ultimate goal to shoot for, but it allows
> comparision.
obviously the current scheduler is not cast into stone, it never was,
never will be.
but determining whether the current behavior is possible in a different
scheduler design is sure a good metric of how flexible that different
scheduler design is.
Ingo
On Wed, 4 Apr 2001, Hubertus Franke wrote:
> Another point to raise is that the current scheduler does a exhaustive
> search for the "best" task to run. It touches every process in the
> runqueue. this is ok if the runqueue length is limited to a very small
> multiple of the #cpus. [...]
indeed. The current scheduler handles UP and SMP systems, up to 32
(perhaps 64) CPUs efficiently. Agressively NUMA systems need a different
approach anyway in many other subsystems too, Kanoj is doing some
scheduler work in that area.
but the original claim was that the scheduling of thousands of runnable
processes (which is not equal to having thousands of sleeping processes)
must perform well - which is a completely different issue.
Ingo
On Wed, Apr 04, 2001 at 03:34:22PM +0200, Ingo Molnar wrote:
>
> On Wed, 4 Apr 2001, Hubertus Franke wrote:
>
> > Another point to raise is that the current scheduler does a exhaustive
> > search for the "best" task to run. It touches every process in the
> > runqueue. this is ok if the runqueue length is limited to a very small
> > multiple of the #cpus. [...]
>
> indeed. The current scheduler handles UP and SMP systems, up to 32
> (perhaps 64) CPUs efficiently. Agressively NUMA systems need a different
> approach anyway in many other subsystems too, Kanoj is doing some
> scheduler work in that area.
I didn't seen anything from Kanoj but I did something myself for the wildfire:
ftp://ftp.us.kernel.org/pub/linux/kernel/people/andrea/kernels/v2.4/2.4.3aa1/10_numa-sched-1
this is mostly an userspace issue, not really intended as a kernel optimization
(however it's also partly a kernel optimization). Basically it splits the load
of the numa machine into per-node load, there can be unbalanced load across the
nodes but fairness is guaranteed inside each node. It's not extremely well
tested but benchmarks were ok and it is at least certainly stable.
However Ingo consider that in a 32-way if you don't have at least 32 tasks
running all the time _always_ you're really stupid paying such big money for
nothing ;). So the fact the scheduler is optimized for 1/2 tasks running all
the time is not nearly enough for those machines (and of course also the
scheduling rate automatically increases linearly with the increase of the
number of cpus). Now it's perfectly fine that we don't ask the embedded and
desktop guys to pay for that, but a kernel configuration option to select an
algorithm that scales would be a good idea IMHO. The above patch just adds a
CONFIG_NUMA_SCHED. The scalable algorithm can fit into it and nobody will be
hurted by that (CONFIG_NUMA_SCHED cannot even be selected by x86 compiles).
Andrea
Hubertus Franke wrote:
>
> This is an important point that Mike is raising and it also addresses a
> critique that Ingo issued yesterday, namely interactivity and fairness.
> The HP scheduler completely separates the per-CPU runqueues and does
> not take preemption goodness or alike into account. This can lead to
> unfair proportionment of CPU cycles, strong priority inversion and a
> potential
> lack of interactivity.
>
> Our MQ scheduler does yield the same decision in most cases
> (other than defined by some race condition on locks and counter members)
>
Let me stress that HP scheduler is not meant to be a replacement for the
current scheduler. The HP scheduler patch allows the current scheduler
to be replaced by another scheduler by loading a module in special
cases. HP is providing three different loadable scheduler modules -
Processor sets, Constant time scheduler, and Multi-runqueue scheduler.
Each one of these is geared towards a specific requirement. I would not
suggest using any of these for a generalized case. Processor sets
scheduler is designed to make scheduling decisions on a per-cpu basis
and not global basis. All we are trying to do is to make the current
scheduler modular so we CAN load an alternate scheduling policy module
in cases where the process mix requires a different scheduling policy or
the site policy require a different scheduling policy. An example of a
specific site processor allocation policy could be a site that runs a
database server on an MP machine along with a few other processes and
the administrator wants to guarantee that the database server process
always gets x percent of processing time or one CPU be dedicated to just
the database server. A policy like this is not meant to be fair and of
course, not a policy we want to impose upon others. The only HP changes
I would put in the kernel sources for general release would be the
changes to scheduler to allow such alternate (not necessarily fair or
the fastest for benchmarks, general process mix or 1000's of processes)
policies to be loaded. When a policy module is not loaded, scheduler
works exactly like the current scheduler even after HP patches. There
are people who could benefit from being able to load alternate policy
schedules. Fabio Ricardi happens to be one of them :-) Anyone who does
not want to even allow an alternate scheduler module to be loaded can
simply compile the alternate scheduler support out and that is how I
would expect most kernels to be compiled, especially the ones that ship
with various distributions. I would like the decision to include support
for alternate scheduler to be made by sys admins themselves for their
individual cases.
--
Khalid
====================================================================
Khalid Aziz Linux Development Laboratory
(970)898-9214 Hewlett-Packard
[email protected] Fort Collins, CO
On Wed, Apr 04, 2001 at 09:44:22AM -0600, Khalid Aziz wrote:
> Let me stress that HP scheduler is not meant to be a replacement for the
> current scheduler. The HP scheduler patch allows the current scheduler
> to be replaced by another scheduler by loading a module in special
> cases.
HP also has a simple mq patch that is _not_ integrated into the pluggable
scheduler framework, I have used it myself.
Christoph
--
Of course it doesn't work. We've performed a software upgrade.
>
>
> On Wed, 4 Apr 2001, Hubertus Franke wrote:
>
> > Another point to raise is that the current scheduler does a exhaustive
> > search for the "best" task to run. It touches every process in the
> > runqueue. this is ok if the runqueue length is limited to a very small
> > multiple of the #cpus. [...]
>
> indeed. The current scheduler handles UP and SMP systems, up to 32
> (perhaps 64) CPUs efficiently. Agressively NUMA systems need a different
> approach anyway in many other subsystems too, Kanoj is doing some
> scheduler work in that area.
Actually, not _much_ work has been done in this area. Alongwith a bunch
of other people, I have some ideas about what needs to be done. For
example, for NUMA, we need to try hard to schedule a thread on the
node that has most of its memory (for no reason other than to decrease
memory latency). Independently, some NUMA machines build in multilevel
caches and local snoops that also means that specific processors on
the same node as the last_processor are also good candidates to run
the process next.
To handle a single layer of shared caches, I have tried certain simple
things, mostly as hacks, but am not pleased with the results yet. More
testing needed.
Kanoj
>
> but the original claim was that the scheduling of thousands of runnable
> processes (which is not equal to having thousands of sleeping processes)
> must perform well - which is a completely different issue.
>
> Ingo
>
>
> _______________________________________________
> Lse-tech mailing list
> [email protected]
> http://lists.sourceforge.net/lists/listinfo/lse-tech
>
>
> I didn't seen anything from Kanoj but I did something myself for the wildfire:
>
> ftp://ftp.us.kernel.org/pub/linux/kernel/people/andrea/kernels/v2.4/2.4.3aa1/10_numa-sched-1
>
> this is mostly an userspace issue, not really intended as a kernel optimization
> (however it's also partly a kernel optimization). Basically it splits the load
> of the numa machine into per-node load, there can be unbalanced load across the
> nodes but fairness is guaranteed inside each node. It's not extremely well
> tested but benchmarks were ok and it is at least certainly stable.
>
Just a quick comment. Andrea, unless your machine has some hardware
that imply pernode runqueues will help (nodelevel caches etc), I fail
to understand how this is helping you ... here's a simple theory though.
If your system is lightly loaded, your pernode queues are actually
implementing some sort of affinity, making sure processes stick to
cpus on nodes where they have allocated most of their memory on. I am
not sure what the situation will be under huge loads though.
As I have mentioned to some people before, percpu/pernode/percpuset/global
runqueues probably all have their advantages and disadvantages, and their
own sweet spots. Wouldn't it be really neat if a system administrator
or performance expert could pick and choose what scheduler behavior he
wants, based on how the system is going to be used?
Kanoj
On Wed, Apr 04, 2001 at 09:39:23AM -0700, Kanoj Sarcar wrote:
> example, for NUMA, we need to try hard to schedule a thread on the
> node that has most of its memory (for no reason other than to decrease
> memory latency). Independently, some NUMA machines build in multilevel
> caches and local snoops that also means that specific processors on
> the same node as the last_processor are also good candidates to run
> the process next.
yes. That will probably need to be optional and choosen by the architecture
at compile time too. The probably most important factor to consider is the
penality of accessing remote memory, I think I can say on all recent and future
machines with a small difference between local and remote memory (and possibly
as you say with a decent cache protocol able to snoop cacheline data from the
other cpus even if they're not dirty) it's much better to always try to keep
the task in its last node. My patch is actually assuming recent machines and it
keeps the task in its last node if not in the last cpu and it keeps doing
memory allocation from there and it forgets about its original node where it
started allocating the memory from. This provided the best performance during
userspace CPU bound load as far I can tell and it also better distribute the load.
Kanoj could you also have a look at the NUMA related common code MM fixes I did
in this patch? I'd like to get them integrated (just skip the arch/alpha/*
include/asm-alpha/* stuff while reading the patch, they're totally orthogonal).
ftp://ftp.us.kernel.org/pub/linux/kernel/people/andrea/kernels/v2.4/2.4.3aa1/00_alpha-numa-1
If you prefer I can extract them in a more finegrinded patch just dropping
the alpha stuff by hand.
Andrea
On Wed, Apr 04, 2001 at 09:50:58AM -0700, Kanoj Sarcar wrote:
> >
> > I didn't seen anything from Kanoj but I did something myself for the wildfire:
> >
> > ftp://ftp.us.kernel.org/pub/linux/kernel/people/andrea/kernels/v2.4/2.4.3aa1/10_numa-sched-1
> >
> > this is mostly an userspace issue, not really intended as a kernel optimization
> > (however it's also partly a kernel optimization). Basically it splits the load
> > of the numa machine into per-node load, there can be unbalanced load across the
> > nodes but fairness is guaranteed inside each node. It's not extremely well
> > tested but benchmarks were ok and it is at least certainly stable.
> >
>
> Just a quick comment. Andrea, unless your machine has some hardware
> that imply pernode runqueues will help (nodelevel caches etc), I fail
> to understand how this is helping you ... here's a simple theory though.
It helps by keeping the task in the same node if it cannot keep it in
the same cpu anymore.
Assume task A is sleeping and it last run on cpu 8 node 2. It gets a wakeup
and it gets running and for some reason cpu 8 is busy and there are other
cpus idle in the system. Now with the current scheduler it can be moved in any
cpu in the system, with the numa sched applied we will try to first reschedule
it in the idles cpus of node 2 for example. The per-node runqueue are mainly
necessary to implement the heuristic.
> cpus on nodes where they have allocated most of their memory on. I am
> not sure what the situation will be under huge loads though.
after all cpus are busy we try to reschedule only on the cpus of the local
node, that's why it can generate some unbalance yes, but it will tend to
rebalance over the time because some node will end with all tasks with
zero counter first if it's less loaded, and so then it will start
getting tasks with has_cpu 0 in the runqueues out of other nodes.
You may want to give it a try on your machines and see what difference it
makes, I'd be curious to know of course.
Andrea
>
> It helps by keeping the task in the same node if it cannot keep it in
> the same cpu anymore.
>
> Assume task A is sleeping and it last run on cpu 8 node 2. It gets a wakeup
> and it gets running and for some reason cpu 8 is busy and there are other
> cpus idle in the system. Now with the current scheduler it can be moved in any
> cpu in the system, with the numa sched applied we will try to first reschedule
> it in the idles cpus of node 2 for example. The per-node runqueue are mainly
> necessary to implement the heuristic.
>
Yes. But this is not the best solution, if I can add on to the example
and make some assumptions.
Imagine that most of the program's memory is on node 1, it was scheduled
on node 2 cpu 8 momentarily (maybe because kswapd ran on node 1, other
higher priority processes took over other cpus on node 1, etc).
Then, your patch will try to keep the process on node 2, which is not
neccessarily the best solution. Of course, as I mentioned before, if
you have a node local cache on node 2, that cache might have been warmed
enough to make scheduling on node 2 a good option.
I am not saying there is a wrong or right answer, there are so many
possibilities, everything probably works and breaks under different
circumstances.
Btw, while we are swapping patches, the patch at
http://oss.sgi.com/projects/numa/download/sched242.patch
tries to implement per-arch scheduling. The current scheduler behavior
is smp_arch_goodness() and smp_pick_cpu(), but the patch allows the
possibility for a specific platform to change that to something else.
Linus has seen this patch, and agrees to it in principle. He does not
consider this 2.4 material though. Of course, I am completely open to
Ingo (or someone else) coming up with a different way of providing the
same freedom to arch specific code.
Kanoj
On Wed, Apr 04, 2001 at 10:49:04AM -0700, Kanoj Sarcar wrote:
> Imagine that most of the program's memory is on node 1, it was scheduled
> on node 2 cpu 8 momentarily (maybe because kswapd ran on node 1, other
> higher priority processes took over other cpus on node 1, etc).
>
> Then, your patch will try to keep the process on node 2, which is not
> neccessarily the best solution. Of course, as I mentioned before, if
> you have a node local cache on node 2, that cache might have been warmed
> enough to make scheduling on node 2 a good option.
Exactly it made it a good option, and more important this heuristic can
only improve performance if compared to the mainline scheduler.
Infact I tried to reschedule the task back to its original node and it dropped
performance, anyways I cannot say to have done an extensive research on that, I
believe if we take care of some more history of the node migration we may
understand it's right time to push it back to its original node but that was
not obvious and I wanted a simple solution to boost the performance under CPU
bound load to start with.
Andrea
Hello
Just dump idea - why not make scheduler switchable with modules - so
users
could select any scheduler they want ?
This should not be that hard and would make it easy to replace scheduler
at runtime so everyone could easily try what's the best for him/her.
[email protected]