Re: [RFT][patch] Scheduling for HTT and not only

On 04/05/12 21:45, Alexander Motin wrote:
> On 05.04.2012 21:12, Arnaud Lacombe wrote:
>> Hi,
>>
>> [Sorry for the delay, I got a bit sidetrack'ed...]
>>
>> 2012/2/17 Alexander Motin<mav_at_freebsd.org>:
>>> On 17.02.2012 18:53, Arnaud Lacombe wrote:
>>>>
>>>> On Fri, Feb 17, 2012 at 11:29 AM, Alexander Motin<mav_at_freebsd.org>
>>>> wrote:
>>>>>
>>>>> On 02/15/12 21:54, Jeff Roberson wrote:
>>>>>>
>>>>>> On Wed, 15 Feb 2012, Alexander Motin wrote:
>>>>>>>
>>>>>>> I've decided to stop those cache black magic practices and focus on
>>>>>>> things that really exist in this world -- SMT and CPU load. I've
>>>>>>> dropped most of cache related things from the patch and made the
>>>>>>> rest
>>>>>>> of things more strict and predictable:
>>>>>>> http://people.freebsd.org/~mav/sched.htt34.patch
>>>>>>
>>>>>>
>>>>>> This looks great. I think there is value in considering the other
>>>>>> approach further but I would like to do this part first. It would be
>>>>>> nice to also add priority as a greater influence in the load
>>>>>> balancing
>>>>>> as well.
>>>>>
>>>>>
>>>>> I haven't got good idea yet about balancing priorities, but I've
>>>>> rewritten
>>>>> balancer itself. As soon as sched_lowest() / sched_highest() are more
>>>>> intelligent now, they allowed to remove topology traversing from the
>>>>> balancer itself. That should fix double-swapping problem, allow to
>>>>> keep
>>>>> some
>>>>> affinity while moving threads and make balancing more fair. I did
>>>>> number
>>>>> of
>>>>> tests running 4, 8, 9 and 16 CPU-bound threads on 8 CPUs. With 4, 8
>>>>> and
>>>>> 16
>>>>> threads everything is stationary as it should. With 9 threads I see
>>>>> regular
>>>>> and random load move between all 8 CPUs. Measurements on 5 minutes run
>>>>> show
>>>>> deviation of only about 5 seconds. It is the same deviation as I see
>>>>> caused
>>>>> by only scheduling of 16 threads on 8 cores without any balancing
>>>>> needed
>>>>> at
>>>>> all. So I believe this code works as it should.
>>>>>
>>>>> Here is the patch: http://people.freebsd.org/~mav/sched.htt40.patch
>>>>>
>>>>> I plan this to be a final patch of this series (more to come :))
>>>>> and if
>>>>> there will be no problems or objections, I am going to commit it
>>>>> (except
>>>>> some debugging KTRs) in about ten days. So now it's a good time for
>>>>> reviews
>>>>> and testing. :)
>>>>>
>>>> is there a place where all the patches are available ?
>>>
>>>
>>> All my scheduler patches are cumulative, so all you need is only the
>>> last
>>> mentioned here sched.htt40.patch.
>>>
>> You may want to have a look to the result I collected in the
>> `runs/freebsd-experiments' branch of:
>>
>> https://github.com/lacombar/hackbench/
>>
>> and compare them with vanilla FreeBSD 9.0 and -CURRENT results
>> available in `runs/freebsd'. On the dual package platform, your patch
>> is not a definite win.
>>
>>> But in some cases, especially for multi-socket systems, to let it
>>> show its
>>> best, you may want to apply additional patch from avg_at_ to better
>>> detect CPU
>>> topology:
>>> https://gitorious.org/~avg/freebsd/avgbsd/commit/6bca4a2e4854ea3fc275946a023db65c483cb9dd
>>>
>>>
>> test I conducted specifically for this patch did not showed much
>> improvement...
>
> If I understand right, this test runs thousands of threads sending and
> receiving data over the pipes. It is quite likely that all CPUs will be
> always busy and so load balancing is not really important in this test,
> What looks good is that more complicated new code is not slower then old
> one.
>
> While this test seems very scheduler-intensive, it may depend on many
> other factors, such as syscall performance, context switch, etc. I'll
> try to play more with it.

My profiling on 8-core Core i7 system shows that code from sched_ule.c 
staying on first places consumes still only 13% of kernel CPU time, 
while doing million of context switches per second. cpu_search(), 
affected by this patch, even less -- only 8%. The rest of time is spread 
between many small other functions. I did some optimizations at r234066 
to reduce cpu_search(0 time to 6%, but looking on how unstable results 
of this test are, hardly any difference there can be really measured by it.

I have strong feeling that while this test may be interesting for 
profiling, it's own results in first place depend not from how fast 
scheduler is, but from the pipes capacity and other alike things. Can 
somebody hint me what except pipe capacity and context switch to 
unblocked receiver prevents sender from sending all data in batch and 
then receiver from receiving them all in batch? If different OSes have 
different policies there, I think results could be incomparable.

-- 
Alexander Motin