[Scst-devel] [ofa-general] SRP/mlx4 interrupts throttling performance

[Vladislav Bolkhovitin]

Cameron Harr, on 02/27/2009 03:18 AM wrote:
> Vladislav Bolkhovitin wrote:
>> Cameron Harr, on 02/26/2009 08:19 PM wrote:
>>> Cameron Harr wrote:
>>>> Cameron Harr wrote:
>>>> I re-compiled and re-ran the tests and numbers are a little better 
>>>> but performance still seems to have gone down from 673:
>>>> Test 1:373751.66
>>>> Test 2:371242.6067
>>>> Test 3:347988.1467
>>>> Test 4:378247.31
>>>> Test 5:375616.53
>>> I was curious and did a regression test with 673 and those numbers 
>>> are now even worse, so I'll presume there is an issue on my system 
>>> and not the SCST code:
>>> Test 1:365204.3067
>>> Test 2:364152.2067
>>> Test 3:340665.7633
>>> Test 4:369916.8133
>>> Test 5:369093.5833
>> It's known that any OS, including Linux, is getting "tired" under load 
>> with time from boot, which leads to worse performance. I guess, you 
>> can experience such effect.
>>
>> Check with r634. R635 has cache locality in data structures related 
>> change, which intended to improve performance a bit, but might make it 
>> worse instead.
>>
> 
> This is with 634. It's pretty bad:
> 338316.44
> 329698.04
> 307972.7133
> 345682.4733
> 344165.08

And 633 is better?

Definitely, you suffer from the system "tiring" effect. So, to get 
comparable results you should do measurements in a predefined state of 
the system, for instance just after boot, and in a row, i.e. one 
immediately after one.

Vlad