Scheduling changes in -mm tree

Scheduling changes in -mm tree

[Martin J. Bligh]

I don't think these are doing much for performance. Or at least 
*something* in your tree isn't ...

Kernbench: 
                                     Elapsed    System      User       CPU
 elm3b67      2.6.11                   50.24    146.60   1117.61   2516.67
 elm3b67      2.6.11-mm1               52.27    141.14   1099.91   2374.33
 elm3b67      2.6.11-mm2               51.88    142.41   1104.85   2403.67
 elm3b67      2.6.11-mm4               51.23    145.04   1100.70   2431.00

(elm3b67 is a 16x x440 ia32 NUMA system + HT)

Is there an easy way to just test those sched changes alone?

M.

Re: Scheduling changes in -mm tree

[Andrew Morton]

"Martin J. Bligh" <mbligh@aracnet.com> wrote:
>
> I don't think these are doing much for performance. Or at least 
> *something* in your tree isn't ...
> 
> Kernbench: 
>                                      Elapsed    System      User       CPU
>  elm3b67      2.6.11                   50.24    146.60   1117.61   2516.67
>  elm3b67      2.6.11-mm1               52.27    141.14   1099.91   2374.33
>  elm3b67      2.6.11-mm2               51.88    142.41   1104.85   2403.67
>  elm3b67      2.6.11-mm4               51.23    145.04   1100.70   2431.00
> 
> (elm3b67 is a 16x x440 ia32 NUMA system + HT)

Sounds like the CPU scheduler, yes

> Is there an easy way to just test those sched changes alone?

Nick has tossed out and redone all the scheduler patches from -mm4, but I
assume it's all pretty much the same.

At http://www.zip.com.au/~akpm/linux/patches/stuff/mbligh.gz is a rollup
(against 2.6.12-rc1) of

sched2-fix-schedstats-warning.patch
sched2-cleanup-wake_idle.patch
sched2-improve-load-balancing-pinned-tasks.patch
sched2-reduce-active-load-balancing.patch
sched2-fix-smt-scheduling-problems.patch
sched2-add-debugging.patch
sched2-less-aggressive-idle-balancing.patch
sched2-balance-timers.patch
sched2-tweak-affine-wakeups.patch
sched2-no-aggressive-idle-balancing.patch
sched2-balance-on-fork.patch
sched2-schedstats-update-for-balance-on-fork.patch
sched2-sched-tuning.patch
sched2-sched-domain-sysctl.patch
add-do_proc_doulonglongvec_minmax-to-sysctl-functions.patch

Re: Scheduling changes in -mm tree

[Martin J. Bligh]

--Andrew Morton <akpm@osdl.org> wrote (on Saturday, March 19, 2005 14:07:54 -0800):

> "Martin J. Bligh" <mbligh@aracnet.com> wrote:
>> 
>> I don't think these are doing much for performance. Or at least 
>> *something* in your tree isn't ...
>> 
>> Kernbench: 
>>                                      Elapsed    System      User       CPU
>>  elm3b67      2.6.11                   50.24    146.60   1117.61   2516.67
>>  elm3b67      2.6.11-mm1               52.27    141.14   1099.91   2374.33
>>  elm3b67      2.6.11-mm2               51.88    142.41   1104.85   2403.67
>>  elm3b67      2.6.11-mm4               51.23    145.04   1100.70   2431.00
>> 
>> (elm3b67 is a 16x x440 ia32 NUMA system + HT)
> 
> Sounds like the CPU scheduler, yes
> 
>> Is there an easy way to just test those sched changes alone?
> 
> Nick has tossed out and redone all the scheduler patches from -mm4, but I
> assume it's all pretty much the same.
> 
> At http://www.zip.com.au/~akpm/linux/patches/stuff/mbligh.gz is a rollup
> (against 2.6.12-rc1) of

Kernbench: 
                                    Elapsed    System      User       CPU
elm3b67      2.6.12-rc1               49.02    147.91   1105.49   2556.00
elm3b67      mbligh                   52.30    142.24   1105.83   2385.33

That doesn't seem like an improvement ;-) (last run is just adding above patch)
I'll try to get you results on a couple more machines, but I'm fighting
with the test harness to get it to behave (plus I now have to rerun all
the tests with CONFIG_BROKEN turned on to get CONFIG_SCSI_QLOGIC_ISP to 
work).

M.