On Tue, Apr 03, 2012 at 11:23:16AM +0100, John Robinson wrote:
> On 02/04/2012 23:48, Jim Kukunas wrote:
> > On Sat, Mar 31, 2012 at 12:38:56PM +0100, John Robinson wrote:
> [...]
> >> I just noticed in my logs the other day (recent el5 kernel on a Core 2):
> >>
> >> raid5: automatically using best checksumming function: generic_sse
> >> generic_sse: 7805.000 MB/sec
> >> raid5: using function: generic_sse (7805.000 MB/sec)
> [...]
> >> raid6: using algorithm sse2x4 (8237 MB/s)
> >>
> >> I was just wondering how it's possible to do the RAID6 Q calculation
> >> faster than the RAID5 XOR calculation - or am I reading this log excerpt
> >> wrongly?
> >
> > Out of curiosity, are you running with CONFIG_PREEMPT=y?
>
> No. Here's an excerpt from my .config:
>
> # CONFIG_PREEMPT_NONE is not set
> CONFIG_PREEMPT_VOLUNTARY=y
> # CONFIG_PREEMPT is not set
> CONFIG_PREEMPT_BKL=y
> CONFIG_PREEMPT_NOTIFIERS=y
>
> But this is a Xen dom0 kernel, 2.6.18-308.1.1.el5.centos.plusxen. Now, a
> non-Xen kernel (2.6.18-308.1.1.el5) says:
> raid5: automatically using best checksumming function: generic_sse
> generic_sse: 11892.000 MB/sec
> raid5: using function: generic_sse (11892.000 MB/sec)
> raid6: int64x1 2644 MB/s
> raid6: int64x2 3238 MB/s
> raid6: int64x4 3011 MB/s
> raid6: int64x8 2503 MB/s
> raid6: sse2x1 5375 MB/s
> raid6: sse2x2 5851 MB/s
> raid6: sse2x4 9136 MB/s
> raid6: using algorithm sse2x4 (9136 MB/s)
>
> Looks like it loses a chunk of performance running as a Xen dom0.
>
> Even still, 11892 MB/s for XOR vs 9136 MB/s for XOR+Q - it still seems
> remarkable that the XOR can't be done several times faster than the Q.
Taking a look at do_xor_speed, I see two issues which might be the cause
of the disparity you reported.
0) In the RAID5 xor benchmark, we get the current jiffy, then run do_2() until
the jiffy increments. This means we could potentially be testing for less
than a full jiffy. The RAID6 benchmark handles this by obtaining the current
jiffy, then calling cpu_relax() until the jiffy increments, and then running
the test. This is addressed by my first patch.
1) The only way I could reproduce your findings of a higher throughput for
RAID6 than for RAID5 xor checksumming was with CONFIG_PREEMPT=y. It seems
that you encountered this while running as XEN dom0. Currently, we disable
preemption during the RAID6 benchmark, but don't in the RAID5 benchmark.
This is addressed by my second patch.
I've added linux-crypto to the discussion as both of these patches affect
code in crypto/
Thanks.
With CONFIG_PREEMPT=y, we need to disable preemption while benchmarking
RAID5 xor checksumming to ensure we're actually measuring what we think
we're measuring.
Signed-off-by: Jim Kukunas <[email protected]>
---
crypto/xor.c | 5 +++++
1 files changed, 5 insertions(+), 0 deletions(-)
diff --git a/crypto/xor.c b/crypto/xor.c
index 8788443..84daa11 100644
--- a/crypto/xor.c
+++ b/crypto/xor.c
@@ -21,6 +21,7 @@
#include <linux/gfp.h>
#include <linux/raid/xor.h>
#include <linux/jiffies.h>
+#include <linux/preempt.h>
#include <asm/xor.h>
/* The xor routines to use. */
@@ -69,6 +70,8 @@ do_xor_speed(struct xor_block_template *tmpl, void *b1, void *b2)
tmpl->next = template_list;
template_list = tmpl;
+ preempt_disable();
+
/*
* Count the number of XORs done during a whole jiffy, and use
* this to calculate the speed of checksumming. We use a 2-page
@@ -91,6 +94,8 @@ do_xor_speed(struct xor_block_template *tmpl, void *b1, void *b2)
max = count;
}
+ preempt_enable();
+
speed = max * (HZ * BENCH_SIZE / 1024);
tmpl->speed = speed;
--
1.7.8.5
In the existing do_xor_speed(), there is no guarantee that we actually
run do_2() for a full jiffy. We get the current jiffy, then run do_2()
until the next jiffy.
Instead, let's get the current jiffy, then wait until the next jiffy
to start our test.
Signed-off-by: Jim Kukunas <[email protected]>
---
crypto/xor.c | 8 +++++---
1 files changed, 5 insertions(+), 3 deletions(-)
diff --git a/crypto/xor.c b/crypto/xor.c
index b75182d..8788443 100644
--- a/crypto/xor.c
+++ b/crypto/xor.c
@@ -63,7 +63,7 @@ static void
do_xor_speed(struct xor_block_template *tmpl, void *b1, void *b2)
{
int speed;
- unsigned long now;
+ unsigned long now, j;
int i, count, max;
tmpl->next = template_list;
@@ -76,9 +76,11 @@ do_xor_speed(struct xor_block_template *tmpl, void *b1, void *b2)
*/
max = 0;
for (i = 0; i < 5; i++) {
- now = jiffies;
+ j = jiffies;
count = 0;
- while (jiffies == now) {
+ while ((now = jiffies) == j)
+ cpu_relax();
+ while (time_before(jiffies, now + 1)) {
mb(); /* prevent loop optimzation */
tmpl->do_2(BENCH_SIZE, b1, b2);
mb();
--
1.7.8.5
[adding Boaz since he also made an attempt at fixing this]
http://marc.info/?l=linux-crypto-vger&m=131829241111450&w=2
...I had meant to follow up on this, but was buried in 'isci' issues.
On Tue, Apr 3, 2012 at 4:56 PM, Jim Kukunas
<[email protected]> wrote:
> On Tue, Apr 03, 2012 at 11:23:16AM +0100, John Robinson wrote:
>> On 02/04/2012 23:48, Jim Kukunas wrote:
>> > On Sat, Mar 31, 2012 at 12:38:56PM +0100, John Robinson wrote:
>> [...]
>> >> I just noticed in my logs the other day (recent el5 kernel on a Core 2):
>> >>
>> >> raid5: automatically using best checksumming function: generic_sse
>> >> ? ? ?generic_sse: ?7805.000 MB/sec
>> >> raid5: using function: generic_sse (7805.000 MB/sec)
>> [...]
>> >> raid6: using algorithm sse2x4 (8237 MB/s)
>> >>
>> >> I was just wondering how it's possible to do the RAID6 Q calculation
>> >> faster than the RAID5 XOR calculation - or am I reading this log excerpt
>> >> wrongly?
>> >
>> > Out of curiosity, are you running with CONFIG_PREEMPT=y?
>>
>> No. Here's an excerpt from my .config:
>>
>> # CONFIG_PREEMPT_NONE is not set
>> CONFIG_PREEMPT_VOLUNTARY=y
>> # CONFIG_PREEMPT is not set
>> CONFIG_PREEMPT_BKL=y
>> CONFIG_PREEMPT_NOTIFIERS=y
>>
>> But this is a Xen dom0 kernel, 2.6.18-308.1.1.el5.centos.plusxen. Now, a
>> non-Xen kernel (2.6.18-308.1.1.el5) says:
>> raid5: automatically using best checksumming function: generic_sse
>> ? ? generic_sse: 11892.000 MB/sec
>> raid5: using function: generic_sse (11892.000 MB/sec)
>> raid6: int64x1 ? 2644 MB/s
>> raid6: int64x2 ? 3238 MB/s
>> raid6: int64x4 ? 3011 MB/s
>> raid6: int64x8 ? 2503 MB/s
>> raid6: sse2x1 ? ?5375 MB/s
>> raid6: sse2x2 ? ?5851 MB/s
>> raid6: sse2x4 ? ?9136 MB/s
>> raid6: using algorithm sse2x4 (9136 MB/s)
>>
>> Looks like it loses a chunk of performance running as a Xen dom0.
>>
>> Even still, 11892 MB/s for XOR vs 9136 MB/s for XOR+Q - it still seems
>> remarkable that the XOR can't be done several times faster than the Q.
>
> Taking a look at do_xor_speed, I see two issues which might be the cause
> of the disparity you reported.
>
> 0) In the RAID5 xor benchmark, we get the current jiffy, then run do_2() until
> the jiffy increments. This means we could potentially be testing for less
> than a full jiffy. The RAID6 benchmark handles this by obtaining the current
> jiffy, then calling cpu_relax() until the jiffy increments, and then running
> the test. This is addressed by my first patch.
>
> 1) The only way I could reproduce your findings of a higher throughput for
> RAID6 than for RAID5 xor checksumming was with CONFIG_PREEMPT=y. It seems
> that you encountered this while running as XEN dom0. Currently, we disable
> preemption during the RAID6 benchmark, but don't in the RAID5 benchmark.
> This is addressed by my second patch.
>
> I've added linux-crypto to the discussion as both of these patches affect
> code in crypto/
>
> Thanks.
>
>
> --
> To unsubscribe from this list: send the line "unsubscribe linux-raid" in
> the body of a message to [email protected]
> More majordomo info at ?http://vger.kernel.org/majordomo-info.html
On 04/06/2012 11:43 PM, Dan Williams wrote:
> [adding Boaz since he also made an attempt at fixing this]
>
> http://marc.info/?l=linux-crypto-vger&m=131829241111450&w=2
>
> ...I had meant to follow up on this, but was buried in 'isci' issues.
>
>
Sorry was traveling.
Yes I have an old fix for this. Which I need to cleanup and retest.
My original problem was an hang in UML, but I noticed the timing problems
as well.
Please give me til the end of the week to settle in and come up to speed.
[Current patch: http://marc.info/?l=linux-crypto-vger&m=131829242311458&w=2]
Thanks
Boaz
> On Tue, Apr 3, 2012 at 4:56 PM, Jim Kukunas
> <[email protected]> wrote:
>> On Tue, Apr 03, 2012 at 11:23:16AM +0100, John Robinson wrote:
>>> On 02/04/2012 23:48, Jim Kukunas wrote:
>>>> On Sat, Mar 31, 2012 at 12:38:56PM +0100, John Robinson wrote:
>>> [...]
>>>>> I just noticed in my logs the other day (recent el5 kernel on a Core 2):
>>>>>
>>>>> raid5: automatically using best checksumming function: generic_sse
>>>>> generic_sse: 7805.000 MB/sec
>>>>> raid5: using function: generic_sse (7805.000 MB/sec)
>>> [...]
>>>>> raid6: using algorithm sse2x4 (8237 MB/s)
>>>>>
>>>>> I was just wondering how it's possible to do the RAID6 Q calculation
>>>>> faster than the RAID5 XOR calculation - or am I reading this log excerpt
>>>>> wrongly?
>>>>
>>>> Out of curiosity, are you running with CONFIG_PREEMPT=y?
>>>
>>> No. Here's an excerpt from my .config:
>>>
>>> # CONFIG_PREEMPT_NONE is not set
>>> CONFIG_PREEMPT_VOLUNTARY=y
>>> # CONFIG_PREEMPT is not set
>>> CONFIG_PREEMPT_BKL=y
>>> CONFIG_PREEMPT_NOTIFIERS=y
>>>
>>> But this is a Xen dom0 kernel, 2.6.18-308.1.1.el5.centos.plusxen. Now, a
>>> non-Xen kernel (2.6.18-308.1.1.el5) says:
>>> raid5: automatically using best checksumming function: generic_sse
>>> generic_sse: 11892.000 MB/sec
>>> raid5: using function: generic_sse (11892.000 MB/sec)
>>> raid6: int64x1 2644 MB/s
>>> raid6: int64x2 3238 MB/s
>>> raid6: int64x4 3011 MB/s
>>> raid6: int64x8 2503 MB/s
>>> raid6: sse2x1 5375 MB/s
>>> raid6: sse2x2 5851 MB/s
>>> raid6: sse2x4 9136 MB/s
>>> raid6: using algorithm sse2x4 (9136 MB/s)
>>>
>>> Looks like it loses a chunk of performance running as a Xen dom0.
>>>
>>> Even still, 11892 MB/s for XOR vs 9136 MB/s for XOR+Q - it still seems
>>> remarkable that the XOR can't be done several times faster than the Q.
>>
>> Taking a look at do_xor_speed, I see two issues which might be the cause
>> of the disparity you reported.
>>
>> 0) In the RAID5 xor benchmark, we get the current jiffy, then run do_2() until
>> the jiffy increments. This means we could potentially be testing for less
>> than a full jiffy. The RAID6 benchmark handles this by obtaining the current
>> jiffy, then calling cpu_relax() until the jiffy increments, and then running
>> the test. This is addressed by my first patch.
>>
>> 1) The only way I could reproduce your findings of a higher throughput for
>> RAID6 than for RAID5 xor checksumming was with CONFIG_PREEMPT=y. It seems
>> that you encountered this while running as XEN dom0. Currently, we disable
>> preemption during the RAID6 benchmark, but don't in the RAID5 benchmark.
>> This is addressed by my second patch.
>>
>> I've added linux-crypto to the discussion as both of these patches affect
>> code in crypto/
>>
>> Thanks.
>>
>>
>> --
>> To unsubscribe from this list: send the line "unsubscribe linux-raid" in
>> the body of a message to [email protected]
>> More majordomo info at http://vger.kernel.org/majordomo-info.html