Re: [RFC][ only for review ] memory controller bacground reclaim [0/5] (Does anyone have an idea about throttling ?)

[Balbir Singh <balbir-23VcF4HTsmIX0ybBhKVfKdBPR1lH4CV8@public.gmane.org>]

KAMEZAWA Hiroyuki wrote:
> On Wed, 28 Nov 2007 17:49:23 +0900
> KAMEZAWA Hiroyuki <kamezawa.hiroyu-+CUm20s59erQFUHtdCDX3A@public.gmane.org> wrote:
> 
>> Hi, this set is for memory controller background reclaim.
>>
>> Merged YAMAMOTO-san's version onto 2.6.23-rc3-mm1 + my NUMA patch.
>> And splitted to several sets.
>>
>> Major changes from his one is
>>  - use kthread instead of work_queue
>>  - adjust high/low watermark when limit changes automatically
>>    and set default value. (a user can specify his own later.)
>>
> FYI rather than RFC.
> 
> I wrote attached patch and run kernbench on 8CPU/2Node NUMA/ia64.
> It does make -j 32.
> 
> Memory limitation was 800M. Low/High watermark here was 750M/780M.
> 
> == These numbers are stable to some extent.==
> 2.6.24-rc3-mm2: (Limit: 800M)     
> Average Optimal -j 32 Load Run:
> Elapsed Time 358.933---------------------------(*)
> User Time 1069.63
> System Time 140.667
> Percent CPU 337.333
> Context Switches 220821
> Sleeps 196912
> 
> 2.6.24-rc3-mm2 + throttle (Limit:800M)
> Average Optimal -j 32 Load Run:
> Elapsed Time 266.697---------------------------(*)
> User Time 1105.39
> System Time 124.423
> Percent CPU 471.667
> Context Switches 251797
> Sleeps 231038
> 
> 2.6.24-rc3-mm2 + throttle + High/Low watermark.
> (low:750M High:780M Limit:800M)
> Average Optimal -j 32 Load Run:
> Elapsed Time 266.844---------------------------(*)
> User Time 1112.9
> System Time 112.273
> Percent CPU 473.667
> Context Switches 251795
> Sleeps 220339
> ==
> 

Looks good to me, was there any impact on memory.failcnt?

> Seems throttling reclaim has some good effect (for kernbench).
> Does anyone have an idea for throttling  reclaiming of memory controller ?
> 

In the past I've run workloads of apache+geronimo+open trade, I've run
linear sequential memory access tests, kernbench, lmbench, database
benchmarks (DOTS, pgbench, etc). I think Lee Schermerhorn has a very
interesting setup (that I need to learn to replicate).

> Thanks,
> -Kame
-- 
	Warm Regards,
	Balbir Singh
	Linux Technology Center
	IBM, ISTL