RE: Why HZ on i386 is 100 ?

[Chen, Kenneth W]

If you change HZ to 1000, you need to change PROC_CHANGE_PENALTY
accordingly.  Otherwise, process would get preempted before its time slice
gets expired.  The net effect is more context switch than necessary, which
could explain the 10% difference.

-----Original Message-----
From: Davide Libenzi [mailto:[e-mail]]
Sent: Tuesday, April 16, 2002 11:10 AM
To: [e-mail]
Cc: Linus Torvalds; Linux Kernel Mailing List
Subject: Re: Why HZ on i386 is 100 ?

On Tue, 16 Apr 2002, David Mosberger wrote:

> >>>>> On Tue, 16 Apr 2002 10:18:18 -0700 (PDT), Davide Libenzi
<[e-mail]> said:
>
>   Davide> i still have pieces of paper on my desk about tests done on
>   Davide> my dual piii where by hacking HZ to 1000 the kernel build
>   Davide> time went from an average of 2min:30sec to an average
>   Davide> 2min:43sec. that is pretty close to 10%
>
> Did you keep the timeslice roughly constant?

it was 2.5.1 time and it was still ruled by TICK_SCALE that made the
timeslice to drop from 60ms ( 100HZ ) to 21ms ( 1000HZ ).

- Davide