Re: [RFC PATCH 0/2] sched: simplify the select_task_rq_fair()

[Michael Wang <wangyun@linux.vnet.ibm.com>]

On 01/21/2013 12:38 PM, Mike Galbraith wrote:
> On Mon, 2013-01-21 at 10:50 +0800, Michael Wang wrote: 
>> On 01/20/2013 12:09 PM, Mike Galbraith wrote:
>>> On Thu, 2013-01-17 at 13:55 +0800, Michael Wang wrote: 
>>>> Hi, Mike
>>>>
>>>> I've send out the v2, which I suppose it will fix the below BUG and
>>>> perform better, please do let me know if it still cause issues on your
>>>> arm7 machine.
>>>
>>> s/arm7/aim7
>>>
>>> Someone swiped half of CPUs/ram, so the box is now 2 10 core nodes vs 4.
>>>
>>> stock scheduler knobs
>>>
>>> 3.8-wang-v2                                 avg     3.8-virgin                          avg    vs wang
>>> Tasks    jobs/min
>>>     1      436.29    435.66    435.97    435.97        437.86    441.69    440.09    439.88      1.008
>>>     5     2361.65   2356.14   2350.66   2356.15       2416.27   2563.45   2374.61   2451.44      1.040
>>>    10     4767.90   4764.15   4779.18   4770.41       4946.94   4832.54   4828.69   4869.39      1.020
>>>    20     9672.79   9703.76   9380.80   9585.78       9634.34   9672.79   9727.13   9678.08      1.009
>>>    40    19162.06  19207.61  19299.36  19223.01      19268.68  19192.40  19056.60  19172.56       .997
>>>    80    37610.55  37465.22  37465.22  37513.66      37263.64  37120.98  37465.22  37283.28       .993
>>>   160    69306.65  69655.17  69257.14  69406.32      69257.14  69306.65  69257.14  69273.64       .998
>>>   320   111512.36 109066.37 111256.45 110611.72     108395.75 107913.19 108335.20 108214.71       .978
>>>   640   142850.83 148483.92 150851.81 147395.52     151974.92 151263.65 151322.67 151520.41      1.027
>>>  1280    52788.89  52706.39  67280.77  57592.01     189931.44 189745.60 189792.02 189823.02      3.295
>>>  2560    75403.91  52905.91  45196.21  57835.34     217368.64 217582.05 217551.54 217500.74      3.760
>>>
>>> sched_latency_ns = 24ms
>>> sched_min_granularity_ns = 8ms
>>> sched_wakeup_granularity_ns = 10ms
>>>
>>> 3.8-wang-v2                                 avg     3.8-virgin                          avg    vs wang
>>> Tasks    jobs/min
>>>     1      436.29    436.60    434.72    435.87        434.41    439.77    438.81    437.66      1.004
>>>     5     2382.08   2393.36   2451.46   2408.96       2451.46   2453.44   2425.94   2443.61      1.014
>>>    10     5029.05   4887.10   5045.80   4987.31       4844.12   4828.69   4844.12   4838.97       .970
>>>    20     9869.71   9734.94   9758.45   9787.70       9513.34   9611.42   9565.90   9563.55       .977
>>>    40    19146.92  19146.92  19192.40  19162.08      18617.51  18603.22  18517.95  18579.56       .969
>>>    80    37177.91  37378.57  37292.31  37282.93      36451.13  36179.10  36233.18  36287.80       .973
>>>   160    70260.87  69109.05  69207.71  69525.87      68281.69  68522.97  68912.58  68572.41       .986
>>>   320   114745.56 113869.64 114474.62 114363.27     114137.73 114137.73 114137.73 114137.73       .998
>>>   640   164338.98 164338.98 164618.00 164431.98     164130.34 164130.34 164130.34 164130.34       .998
>>>  1280   209473.40 209134.54 209473.40 209360.44     210040.62 210040.62 210097.51 210059.58      1.003
>>>  2560   242703.38 242627.46 242779.34 242703.39     244001.26 243847.85 243732.91 243860.67      1.004
>>>
>>> As you can see, the load collapsed at the high load end with stock
>>> scheduler knobs (desktop latency).  With knobs set to scale, the delta
>>> disappeared.
>>
>> Thanks for the testing, Mike, please allow me to ask few questions.
>>
>> What are those tasks actually doing? what's the workload?
> 
> It's the canned aim7 compute load, mixed bag load weighted toward
> compute.  Below is the workfile, should give you an idea.
> 
> # @(#) workfile.compute:1.3 1/22/96 00:00:00
> # Compute Server Mix
> FILESIZE: 100K
> POOLSIZE: 250M
> 50  add_double
> 30  add_int
> 30  add_long
> 10  array_rtns
> 10  disk_cp
> 30  disk_rd
> 10  disk_src
> 20  disk_wrt
> 40  div_double
> 30  div_int
> 50  matrix_rtns
> 40  mem_rtns_1
> 40  mem_rtns_2
> 50  mul_double
> 30  mul_int
> 30  mul_long
> 40  new_raph
> 40  num_rtns_1
> 50  page_test
> 40  series_1
> 10  shared_memory
> 30  sieve
> 20  stream_pipe
> 30  string_rtns
> 40  trig_rtns
> 20  udp_test
> 

That seems like the default one, could you please show me the numbers in
your datapoint file?

I'm not familiar with this benchmark, but I'd like to have a try on my
server, to make sure whether it is a generic issue.

>> And I'm confusing about how those new parameter value was figured out
>> and how could them help solve the possible issue?
> 
> Oh, that's easy.  I set sched_min_granularity_ns such that last_buddy
> kicks in when a third task arrives on a runqueue, and set
> sched_wakeup_granularity_ns near minimum that still allows wakeup
> preemption to occur.  Combined effect is reduced over-scheduling.

That sounds very hard, to catch the timing, whatever, it could be an
important clue for analysis.

>> Do you have any idea about which part in this patch set may cause the issue?
> 
> Nope, I'm as puzzled by that as you are.  When the box had 40 cores,
> both virgin and patched showed over-scheduling effects, but not like
> this.  With 20 cores, symptoms changed in a most puzzling way, and I
> don't see how you'd be directly responsible.

Hmm...

> 
>> One change by designed is that, for old logical, if it's a wake up and
>> we found affine sd, the select func will never go into the balance path,
>> but the new logical will, in some cases, do you think this could be a
>> problem?
> 
> Since it's the high load end, where looking for an idle core is most
> likely to be a waste of time, it makes sense that entering the balance
> path would hurt _some_, it isn't free.. except for twiddling preemption
> knobs making the collapse just go away.  We're still going to enter that
> path if all cores are busy, no matter how I twiddle those knobs.

May be we could try change this back to the old way later, after the aim
7 test on my server.

>   
>>> I thought perhaps the bogus (shouldn't exist) CPU domain in mainline
>>> somehow contributes to the strange behavioral delta, but killing it made
>>> zero difference.  All of these numbers for both trees were logged with
>>> the below applies, but as noted, it changed nothing. 
>>
>> The patch set was supposed to do accelerate by reduce the cost of
>> select_task_rq(), so it should be harmless for all the conditions.
> 
> Yeah, it should just save some cycles, but I like to eliminate known
> bugs when testing, just in case.

Agree, that's really important.

Regards,
Michael Wang

> 
> -Mike
>