Re: [ck] CPU load balance or one task domination under linux-2.6.7-ruby-s77

[Con K. <ke...@ko...>]

Aivils wrote:
> On Monday 05 July 2004 21:19, Con Kolivas wrote:
> 
>>Aivils wrote:
>>
>>>Hi All!
>>>
>>>	I try multiple local user linux box.
>>>All 2.6.XX kernel versions from 2.6.0 up to 2.6.7 do one task domination.
>>>I applay before staircase 7 and now staircase 7.7
>>>
>>>Anyway if two users run games simultaneous, then one game allways
>>>dominate.
>>>If i set "nice -n +1 game-bin"  for each user , then both games
>>>became choppy, because lost priority against system tasks.
>>>
>>>Second phenomen - tuxracer under 2.4.26 take 1%-3% CPU time,
>>>but under 2.6.7-s77 it eat 80%. As result i cannot test thiny games
>>>simultaneous, because thiny goes fat and choppy.
>>>
>>>Does know anybody is this phenomen realy depended from task
>>>scheduler alghoritm?
>>>
>>>May be that is system call queue?
>>
>>Hi
>>
>>Please try staircase7.8 and use
>>echo 0 > /proc/sys/kernel/interactive
> 
> 
> Ok. I was try two kernels 2.6.7,
> 1) with vanilla scheduler non-preemptive,
> 2) staircase 7.8 w/ echo 0 > /proc/sys/kernel/interactive
> both have 100Hz kernel timer
> Two users start game
> otto> nice -n +10 game-bin
> fritz> nice -n +10 game-bin
> 
> Gamers feeling under both kernels are identical.

That's a good start.

> Kernel normal split resources almost righteous for each game.
> BUT THIS IS NOT 2.4.XX FEELING!

Well 2.4 was not an O(1) kernel so with the good comes the bad in 2.6

> 
> Seems  2.6.XX rescheduling time is larger than magic 10-30m sec
> which gamer cannot feel visual.

Actually rescheduling on staircase is in the order of 10ms.

> Under 2.4.XX on second game start fisrt gamer feel on frames per
> second (fps) decrease, but game still smooth.
> Under 2.6.XX on second game start fisrt gamer feel horrible
> delays. Almost game become unplayable.
> 
> Under 2.6.XX is impossible start two simultaneous games without
> nice +

Try both at nice +19.

>>cpu time accounting can be out by a great margin changing Hz (as happens 
>>between 2.4 and 2.6). Higher Hz will always show higher cpu usage even 
>>if the usage is the same. If a task waes up at just the right time 
>>frequently enough it can appear to use 10 times as much cpu.
> 
> 100 Hz does the work :)
> tuxracer is back and take 1% CPU as under 2.4.XX

This doesn't mean it's using less cpu; it means it's reporting less cpu.

> 
> Me never  trust lest usage is the same as under 2.4.XX, because two
> tuxracers under 2.6.7 distrub each other!
> 
> Aivils

Thanks for feedback.

Cheers,
Con