| 1 |
On Thursday 15 May 2008, Mick wrote: |
| 2 |
> > You'll probably get better results with X by selecting a suitable |
| 3 |
> > process scheduler and configuring your HZ to 1000 |
| 4 |
> |
| 5 |
> Now, this I have noticed making a difference. Not all schedulers are |
| 6 |
> born the same. I have found that (the current version of) CFQ is |
| 7 |
> better than others. |
| 8 |
> |
| 9 |
> As a matter of interest, I remember reading somewhere that squeezing |
| 10 |
> 1000Hz out of an old machine may have the opposite effect to that |
| 11 |
> intended. Is this pub talk, or have you experienced something that |
| 12 |
> confirms this? |
| 13 |
|
| 14 |
No, it's not just pub talk. The trick is to look closely at what is |
| 15 |
happening and why. |
| 16 |
|
| 17 |
The HZ value indicates how often the kernel should "tick", which is a |
| 18 |
timing signal. The tick consumes resources of course, but has the |
| 19 |
benefit of accurate timing signals. Modern machines can cope with this |
| 20 |
nicely, they are fast enough. Older machines, in combination with the |
| 21 |
kind of software we run these days, can't cope with this amount of |
| 22 |
activity, and the whole system slows down. The problem is very dynamic |
| 23 |
and subject to many variables so there is no single one-shot solution. |
| 24 |
The answer to what to do very much starts with "It depends" |
| 25 |
|
| 26 |
This is why Con first started working on process schedulers, looking for |
| 27 |
an algorithm the kernel could use to adapt to these cases and still be |
| 28 |
responsive on the desktop. The man with the numbers to back it all up |
| 29 |
is Ingo Molnar. You can read the lkml archives from a few months back |
| 30 |
when CFQ was going through heavy development to get an idea of how |
| 31 |
tricky this can really get |
| 32 |
|
| 33 |
-- |
| 34 |
Alan McKinnon |
| 35 |
alan dot mckinnon at gmail dot com |
| 36 |
|
| 37 |
-- |
| 38 |
gentoo-user@l.g.o mailing list |
Archives