Re: [patch 0/9] sched: use runnable load avg in balance

[Alex Shi <alex.shi@intel.com>]

On 06/07/2013 03:20 PM, Alex Shi wrote:
> Peter&Ingo:
> 
> v8: add a marco div64_ul and used it in task_h_load()
> v7: rebasing on tip/sched/core tree.

Peter & Ingo:

IMHO, if the patch set missed in 3.11 kernel. We will lose the following
widely benefit on many benchmarks, hackbench, pgbench, sysbench,
anonymous java load ...

What's your opinions? :)

> 
> I tested on Intel core2, NHM, SNB, IVB, 2 and 4 sockets machines with
> benchmark kbuild, aim7, dbench, tbench, hackbench, oltp, and netperf
> loopback etc. 
> 
> On SNB EP 4 sockets machine, the hackbench increased about 50%, and
> result become stable. on other machines, hackbench increased about
> 2~10%. oltp increased about 30% in NHM EX box. netperf loopback also
> increased on SNB EP 4 sockets box. 
> No clear changes on other benchmarks.
> 
> Michael Wang gotten better performance on pgbench on his box with this
> patchset. https://lkml.org/lkml/2013/5/16/82
> 
> And Morten tested previous version with better power consumption.
> http://comments.gmane.org/gmane.linux.kernel/1463371
> 
> Changlong found ltp cgroup stress testing get faster on SNB EP
> machine with the last patch.  https://lkml.org/lkml/2013/5/23/65
> ---
> 3.10-rc1          patch1-7         patch1-8
> duration=764   duration=754   duration=750
> duration=764   duration=754   duration=751
> duration=763   duration=755   duration=751
> 
> duration means the seconds of testing cost.
> ---
> 
> Jason also found java server load benefited on his 8 sockets machine
> with last patch. https://lkml.org/lkml/2013/5/29/673
> ---
> When using a 3.10-rc2 tip kernel with patches 1-8, there was about a 40%
> improvement in performance of the workload compared to when using the
> vanilla 3.10-rc2 tip kernel with no patches. When using a 3.10-rc2 tip
> kernel with just patches 1-7, the performance improvement of the
> workload over the vanilla 3.10-rc2 tip kernel was about 25%.
> ---
> 
> We also tried to include blocked load avg in balance. but find many
> benchmark performance drop a lot! Seems accumulating current
> blocked_load_avg into cpu load isn't a good idea. Because:
> 1, The blocked_load_avg is decayed same as runnable load, sometime is far
> bigger than runnable load, that drive tasks to other idle or slight
> load cpu, than cause both performance and power issue. But if the
> blocked load is decayed too fast, it lose its effect. 
> 2, Another issue of blocked load is that when waking up task, we can not 
> know blocked load proportion of the task on rq. So, the blocked load is
> meaningless in wake affine decision.  
> 
> According to above problem, we can not figure out a right way now to use 
> blocked_load_avg in balance.
> 
> Since using runnable load avg in balance brings much benefit on
> performance and power. and this patch was reviewed for long time.
> So seems it's time to let it be clobbered in some sub tree, like 
> tip or linux-next.  Any comments?
> 
> Regards
> Alex
> 


-- 
Thanks
    Alex