Re: Inconsistent load average on tickless kernels

["Lesław Kopeć" <leslaw.kopec@nasza-klasa.pl>]

[-- Attachment #1: Type: text/plain, Size: 1424 bytes --]

Hello!

I've finally finished testing patch c308b56b53. Apologies for the delay
in reporting back. This time I've compared kernels from 3.2 and 2.6.32
branches. Here's a snapshot of load 15 on different versions:

2.6.32.55-hz-0f004f5a69			12.78
2.6.32.55-no-hz-74f5187ac8		4.42
2.6.32.55-no-hz-0f004f5a69		0.49
2.6.32.55-no-hz-c308b56b53		4.37
3.2.12-hz				12.85
3.2.12-no-hz				0.65
3.2.12-no-hz-c308b56b53			7.25
3.2.12-no-hz-c308b56b53	nohz=off	10.59

A whole day trend is available on a comparison chart [1].

Just to make things clear 2.6.32 kernels were patched as follows:
* 74f5187ac8 - just 74f5187ac8
* 0f004f5a69 - 74f5187ac8 + 0f004f5a69
* c308b56b53 - 74f5187ac8 + 0f004f5a69 + c308b56b53

For kernel 3.2.12 patch c308b56b53 seems almost perfect. For low CPU
utilization the load value is slightly lower for NO_HZ version than it
is for HZ one. However the difference is small and the overall trend
relates to CPU usage quite closely. This is definitely the best match so
far. Thanks!

Looking at results for 2.6.32.55 branch it seems that we're back at
74f5187ac8 patch - the values are almost the same. The difference
between NO_HZ and HZ versions is noticeable. At the risk of sounding
like an ungrateful bastard - will there be further attempts at fixing
this bug for 2.6.32 kernels?


[1] http://img535.imageshack.us/img535/2204/kernelload.png

-- 
Lesław Kopeć


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 262 bytes --]