Re: [Lse-tech] VolanMark results for 2.4.4+MQ scheduler

[Bill H. <ha...@au...>]

Gerrit Huizenga wrote:
> 
> So if I read this correctly, this suggests that if I buy 8 CPUs for my
> machine, I used to get the equivalent of one and one third CPUs worth of
> throughput on my machine, and now I'm up to getting the equivalent of
> three and a half CPUs worth of work down with this benchmark?

At the moment this is true.

> 
> So this is definitely dramatic improvement, but what keeps this
> benchmark from getting me the equivalent of, say, 7.5 or 8 CPUs
> worth of work out of this benchmark?  Obviously the scheduler is
> in the way - what's the next bottleneck visible on the performance
> highway?

VolanoMark is capable of scaling 7+ / 8 on at least a couple of
different operating systems.

The MQ scheduler eliminated the contention on the runqueue lock.  This
is apparent from lockmeter where we now have no locks with significant
contention.  I will post the lockmeter for 2.4.4, 2.4.4+MQ.  I will also
post kernprof acg from 2.4.4 and 2.4.4+MQ.

Here is part of the acg for 2.4.4+MQ on an 8-way :

Each sample counts as 0.01 seconds.
  %   cumulative   self              self     total
 time   seconds   seconds    calls  us/call  us/call  name
 27.26     32.25    32.25                             cg_record_arc
 24.75     61.54    29.29                             USER
 11.96     75.69    14.15    20786   680.75   680.75  poll_idle
  7.42     84.47     8.78  1429700     6.14     6.35  schedule
  2.42     87.33     2.86   314872     9.08     9.32  send_sig_info
  1.70     89.34     2.01   313515     6.41     6.41  _wake_up
  1.42     91.02     1.68   314871     5.34     8.05  do_signal
  0.65    100.51     0.77   859520     0.90     0.90  sys_sched_yield

So if you eliminate cg_record_arc, USER, poll_idle and look at
the remaining time spent in the kernel :

100-27.26-24.75-11.96 = 36.03.
Of the 36.03%, (7.42+1.70+0.65)/36.03 = 27.1 % is still scheduler related.

I think further analysis of the MQ scheduler is needed.

> 
> gerrit
> 
> > Throughput results reported are messages/second.
> > Scaling results are relative to UP.
> >
> >          2.4.4     2.4.4      2.4.4+MQ
> >          8M/64M    64M/64M    64M/64M
> > Throughput:
> > UP       12593     13123      13123*
> > 4P       19748     22331      29440
> > 8P       16693     18677      46166
> > Scaling :
> > 4P       1.57      1.70       2.24
> > 8P       1.33      1.42       3.52
> 
> _______________________________________________
> Lse-tech mailing list
> Lse...@li...
> http://lists.sourceforge.net/lists/listinfo/lse-tech