Re: aim7 -30% regression in 2.6.24-rc1

[Cyrus Massoumi <cyrusm@gmx.net>]

Zhang, Yanmin wrote:
> On Thu, 2007-11-01 at 11:02 +0100, Cyrus Massoumi wrote:
>> Zhang, Yanmin wrote:
>>> On Wed, 2007-10-31 at 17:57 +0800, Zhang, Yanmin wrote:
>>>> On Tue, 2007-10-30 at 16:36 +0800, Zhang, Yanmin wrote:
>>>>> On Tue, 2007-10-30 at 08:26 +0100, Ingo Molnar wrote:
>>>>>> * Zhang, Yanmin <yanmin_zhang@linux.intel.com> wrote:
>>>>>>
>>>>>>> sub-bisecting captured patch 
>>>>>>> 38ad464d410dadceda1563f36bdb0be7fe4c8938(sched: uniform tunings) 
>>>>>>> caused 20% regression of aim7.
>>>>>>>
>>>>>>> The last 10% should be also related to sched parameters, such like 
>>>>>>> sysctl_sched_min_granularity.
>>>>>> ah, interesting. Since you have CONFIG_SCHED_DEBUG enabled, could you 
>>>>>> please try to figure out what the best value for 
>>>>>> /proc/sys/kernel_sched_latency, /proc/sys/kernel_sched_nr_latency and 
>>>>>> /proc/sys/kernel_sched_min_granularity is?
>>>>>>
>>>>>> there's a tuning constraint for kernel_sched_nr_latency: 
>>>>>>
>>>>>> - kernel_sched_nr_latency should always be set to 
>>>>>>   kernel_sched_latency/kernel_sched_min_granularity. (it's not a free 
>>>>>>   tunable)
>>>>>>
>>>>>> i suspect a good approach would be to double the value of 
>>>>>> kernel_sched_latency and kernel_sched_nr_latency in each tuning 
>>>>>> iteration, while keeping kernel_sched_min_granularity unchanged. That 
>>>>>> will excercise the tuning values of the 2.6.23 kernel as well.
>>>>> I followed your idea to test 2.6.24-rc1. The improvement is slow.
>>>>> When sched_nr_latency=2560 and sched_latency_ns=640000000, the performance
>>>>> is still about 15% less than 2.6.23.
>>>> I got the aim7 30% regression on my new upgraded stoakley machine. I found
>>>> this mahcine is slower than the old one. Maybe BIOS has issues, or memeory(Might not
>>>> be dual-channel?) is slow. So I retested it on the old machine and found on the old
>>>> stoakley machine, the regression is about 6%, quite similiar to the regression on tigerton
>>>> machine.
>>>>
>>>> By sched_nr_latency=640 and sched_latency_ns=640000000 on the old stoakley machine,
>>>> the regression becomes about 2%. Other latency has more regression.
>>>>
>>>> On my tulsa machine, by sched_nr_latency=640 and sched_latency_ns=640000000,
>>>> the regression becomes less than 1% (The original regression is about 20%).
>>> I rerun SPECjbb by ched_nr_latency=640 and sched_latency_ns=640000000. On tigerton,
>>> the regression is still more than 40%. On stoakley machine, it becomes worse (26%,
>>> original is 9%). I will do more investigation to make sure SPECjbb regression is
>>> also casued by the bad default values.
>>>
>>> We need a smarter method to calculate the best default values for the key tuning
>>> parameters.
>>>
>>> One interesting is sysbench+mysql(readonly) got the same result like 2.6.22 (no
>>> regression). Good job!
>> Do you mean you couldn't reproduce the regression which was reported 
>> with 2.6.23 (http://lkml.org/lkml/2007/10/30/53) with 2.6.24-rc1?
> It looks like you missed my emails.

Yeah :(

> Firstly, I reproduced (or just find the same myself :) ) the issue with kernel 2.6.22,
> 2.6.23-rc and 2.6.23.
> 
> Ingo wrote a big patch to fix it and the new patch is in 2.6.24-rc1 now.

That's nice, could you please point me to the commit?

> Then I retested it with 2.6.24-rc1 on a couple of x86_64 machines. The issue
> disappeared. You could test it with 2.6.24-rc1.

Will do!

>>  It 
>> would be nice if you could provide some numbers for 2.6.22, 2.6.23 and 
>> 2.6.24-rc1.
> Sorry. Intel policy doesn't allow me to publish the numbers because only
> specific departments in Intel could do that. But I could talk the regression
> percentage.

Fair enough :)

> -yanmin

greetings
Cyrus