Re: [PATCH] sched: wakeup buddy

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

[snip]
>> It could bring the same benefit but at lower overhead, what's the point
>> of computing the same value over and over again? Also, the rate limit
>> thing naturally works for the soft/hard-irq case.
> 
> Just try to confirm my understanding, so we are going to do something
> like:
> 
> 	if (now - wakee->last > time_limit) && wakeup_affine()
> 		wakee->last = now
> 		select_idle_sibling(curr_cpu)
> 	else
> 		select_idle_sibling(prev_cpu)
> 
> And time_limit is some static value respect to the rate of load balance,
> is that correct?
> 
> Currently I haven't found regression by reduce the rate, but if we found
> such benchmark, we may still need a way (knob or CONFIG) to disable this
> limitation.

I've done some fast tests on this proposal, on my 12 cpu box, the
pgbench 32 clients test, for a 1000ms time_limit, the benefit is just
like the 8 ref wakeup buddy, when adopt 10ms time_limit, the benefit
dropped half, when time_limit is 1ms, the benefit is less than 10%.

			tps

original		43404

wakeup-buddy		63024		+45.20%

1s-limit		62359		+43.67%
100ms-limit		57547		+32.58%
10ms-limit		52258		+20.40%
1ms-limit		46535		+7.21%

Other test items of pgbench are corresponding, and other benchmarks
still inert to the changes.

I'm planning to make a new patch for this approach later, in which
time_limit is a knob with the default value 1ms (usually the initial
value of balance_interval and the value of min_interval), that will
based on the latest tip tree.

Regards,
Michael Wang