RE: [PATCH] sched: Folding nohz load accounting more accurate

["Doug Smythies" <dsmythies@telus.net>]

[-- Attachment #1: Type: text/plain, Size: 1611 bytes --]

>  On 2012.06.12 02:56 - 0800 (I think), Peter Zijlstra wrote:

>Also added Doug to CC, hopefully we now have everybody who pokes at this
>stuff.

Thanks.

On my computer, and from a different thread from yesterday, I let
the proposed "wang" patch multiple processes test continue for
another 24 hours. The png file showing the results is attached, also
available at [1].

Conclusion: The proposed "wang" patch is worse for the lower load
conditions, giving higher reported load average errors for the same
conditions. The proposed "wang" patch tends towards a load equal to
the number of processes, independent of the actual load of those
processes.

Interestingly, with the "wang" patch I was able to remove the 10
tick grace period without bad side effects (very minimally tested).

@ Charles or Tao: If I could ask: What is your expected load for your 16
processes case? Because you used to get a reported load average of
< 1, we know that the processes enter and exit idle (sleep) at a high
frequency (as that was only possible way for the older under reporting
issue, at least as far as I know). You said it now reports a load
average of 8 to 10, but that is too low. How many CPU's do you have?
I have been unable to re-create your situation on my test computer
(an i7 CPU).
When I run 16 processes, where each process would use 0.95 of a cpu,
if the system did not become resource limited, I get a reported load
average of about 15 to 16. Kernel = 3.5 RC2. Process sleep frequency
was about 80 Hertz each.

[1]
http://www.smythies.com/~doug/network/load_average/load_processes_wang.html

Doug Smythies


[-- Attachment #2: load_processes_wang.png --]
[-- Type: image/png, Size: 38927 bytes --]