Re: [PATCH v2] sched: wake-affine throttle
From: Michael Wang
Date: Tue May 28 2013 - 01:02:45 EST
On 05/22/2013 10:55 PM, Mike Galbraith wrote:
> On Wed, 2013-05-22 at 17:25 +0800, Michael Wang wrote:
>
>> I've not test the hackbench with wakeup-buddy before, will do it this
>> time, I suppose the 15% illegal income will suffered, anyway, it's
>> illegal :)
>
> On a 4 socket 40 core (+SMT) box, hackbench wasn't too happy.
I've done more test and now I got the reason of regression...
The writer and reader in hackbench is N:N, prev writer will write all
the fd then switch to next writer and repeat the same work, so it's
impossible to setup the buddy relationship by just record the last one,
and we have to record all the waker/wakee in history, but that means
unacceptable memory overhead...
So this buddy idea seems to be bad...
I think a better way may should be allowing pull in most time, but
filter the very bad cases carefully.
For workload like pgbench, we actually just need to avoid pull if that
will damage the 'mother' thread, which is busy and be relied by many
'child'.
I will send out a new implementation, let's see whether it could solve
the problems ;-)
Regards,
Michael Wang
>
> defaultx = wakeup buddy
>
> tbench 30 localhost
> 1 5 10 20 40 80 160 320
> 3.10.0-default 188.13 953.03 1860.03 3180.93 5378.34 10826.06 11342.90 11651.30
> 3.10.0-defaultx 187.26 934.55 1859.30 3160.29 5016.35 12477.15 12567.05 12068.47
>
> hackbench -l 400 -g 400
> 3.10.0-default Time: 3.919 4.250 4.116
> 3.10.0-defaultx Time: 9.074 10.985 9.849
>
> aim7 compute
> AIM Multiuser Benchmark - Suite VII "1.1" 3.10.0-default AIM Multiuser Benchmark - Suite VII "1.1" 3.10.0-defaultx
>
> Tasks Jobs/Min JTI Real CPU Jobs/sec/task Tasks Jobs/Min JTI Real CPU Jobs/sec/task
> 1 428.0 100 14.2 4.1 7.1328 1 428.9 100 14.1 4.1 7.1479
> 1 417.9 100 14.5 4.1 6.9655 1 430.4 100 14.1 4.0 7.1733
> 1 427.7 100 14.2 4.2 7.1277 1 424.7 100 14.3 4.2 7.0778
> 5 2350.7 99 12.9 13.8 7.8355 5 2156.6 99 14.1 19.8 7.1886
> 5 2422.1 99 12.5 12.1 8.0735 5 2155.0 99 14.1 19.7 7.1835
> 5 2189.3 98 13.8 18.3 7.2977 5 2108.6 99 14.4 21.4 7.0285
> 10 4515.6 93 13.4 27.6 7.5261 10 4529.1 96 13.4 29.5 7.5486
> 10 4708.6 96 12.9 24.3 7.8477 10 4597.9 96 13.2 26.9 7.6631
> 10 4636.6 96 13.1 25.7 7.7276 10 5197.3 98 11.7 14.8 8.6621
> 20 8053.2 95 15.1 78.1 6.7110 20 9431.9 98 12.8 49.1 7.8599
> 20 8250.5 92 14.7 67.5 6.8754 20 7973.7 97 15.2 93.4 6.6447
> 20 8178.1 97 14.8 78.5 6.8151 20 8145.2 95 14.9 78.4 6.7876
> 40 17413.8 94 13.9 88.6 7.2557 40 16312.2 92 14.9 115.6 6.7968
> 40 16775.1 93 14.5 111.6 6.9896 40 17070.4 94 14.2 110.6 7.1127
> 40 16031.7 93 15.1 147.1 6.6799 40 17578.0 94 13.8 96.9 7.3241
> 80 33666.7 95 14.4 138.4 7.0139 80 33854.7 96 14.3 177.9 7.0531
> 80 33949.6 97 14.3 128.0 7.0728 80 34164.9 96 14.2 146.4 7.1177
> 80 35752.2 96 13.6 159.1 7.4484 80 33807.5 96 14.3 127.6 7.0432
> 160 74814.8 98 13.0 149.8 7.7932 160 75162.8 98 12.9 148.6 7.8295
> 160 74015.3 97 13.1 149.5 7.7099 160 74642.0 98 13.0 168.2 7.7752
> 160 73621.9 98 13.2 146.3 7.6689 160 75572.9 98 12.8 163.6 7.8722
> 320 139210.3 96 13.9 280.1 7.2505 320 139010.8 97 14.0 282.4 7.2401
> 320 135135.9 96 14.3 277.8 7.0383 320 139611.2 96 13.9 282.2 7.2714
> 320 139110.5 96 13.9 280.4 7.2453 320 138514.3 97 14.0 281.0 7.2143
> 640 223538.9 98 17.4 577.2 5.8213 640 224704.5 95 17.3 567.3 5.8517
> 640 224055.5 97 17.3 575.7 5.8348 640 222385.3 95 17.4 580.1 5.7913
> 640 225488.4 96 17.2 566.3 5.8721 640 225882.4 93 17.2 563.0 5.8824
>
>
> --
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to majordomo@xxxxxxxxxxxxxxx
> More majordomo info at http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at http://www.tux.org/lkml/
>
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/