> That would be nice... but no -
>
> both -aa and -ac have it though, and it seems
> solid, so maybe there's hope for 2.4 mainline
> getting it eventually.
>
> Joe
>
>
It seems that the version of O(1) scheduler on 2.4.19-pre10-ac2 is not
perfect (see below), but I asked because it gave me overall performance
gains, specially in multithreading programs (and now I'm going to try
with ngpt 2.00). At least that is the first impression, I'm trying it
for a few days.
I said "not perfect" because a rather non-important benchmarking called
quake 3 seens a lot worse in pre10-ac2 with preemptive patches when
compared against -pre10 with preemptive patches: sound and screen popped
sometimes, like if there was a background task borrowing some cpu, which
was not the case, I mean, no other background tasks compared with
testing against -pre10. That was the only exception to the above
paragraph that I can remember of.
Anyway, that "benchmarking tool" isn't rather serious and I'm no
benchmarking expert :-)
Thank you,
Alexandre
Alexandre P. Nunes wrote:
>> That would be nice... but no -
>>
>> both -aa and -ac have it though, and it seems
>> solid, so maybe there's hope for 2.4 mainline
>> getting it eventually.
>>
>> Joe
>>
>>
>
> I said "not perfect" because a rather non-important benchmarking
> called quake 3 seens a lot worse in pre10-ac2 with preemptive patches
> when compared against -pre10 with preemptive patches: sound and screen
> popped sometimes, like if there was a background task borrowing some
> cpu, which was not the case, I mean, no other background tasks
> compared with testing against -pre10. That was the only exception to
> the above paragraph that I can remember of.
I am running -pre10-ac2 on one box - but also
the preemptive patch - are you running that?
On another box, I'm running -pre10aa4, which
also includes the low latency mini patch.
I also follow the q3a benchmark closely, and
RtCW as well. (The enemy is weakened!)
-ac + preempt and -aa both seem OK here, but
I prefer -aa since I use tux.
Joe
[email protected]> <[email protected]> <1
025128477.1144.3.camel@icbm>
<[email protected]> <1025192465.1084.3.camel@icbm>
<[email protected]> <[email protected]>
<1025202738.1084.12.camel@icbm> <[email protected]>
<[email protected]> <[email protected]>
Content-Type: text/plain
Content-Transfer-Encoding: 7bit
X-Mailer: Ximian Evolution 1.0.3 (1.0.3-6)
Date: 27 Jun 2002 16:35:46 -0400
Message-Id: <1025210179.1080.22.camel@icbm>
Mime-Version: 1.0
On Thu, 2002-06-27 at 16:23, Alexandre P. Nunes wrote:
> It seems that the version of O(1) scheduler on 2.4.19-pre10-ac2 is not
> perfect (see below), but I asked because it gave me overall performance
> gains, specially in multithreading programs (and now I'm going to try
> with ngpt 2.00). At least that is the first impression, I'm trying it
> for a few days.
Alan has some patches queued and I will continue to send him updates as
we get them into 2.5 and they prove stable.
I also will update my 2.4 O(1) scheduler patches when I return from
OLS. This would allow a 2.4-ac vs 2.4-O(1) test.
> I said "not perfect" because a rather non-important benchmarking called
> quake 3 seens a lot worse in pre10-ac2 with preemptive patches when
> compared against -pre10 with preemptive patches: sound and screen popped
> sometimes, like if there was a background task borrowing some cpu, which
> was not the case, I mean, no other background tasks compared with
> testing against -pre10. That was the only exception to the above
> paragraph that I can remember of.
There is some "rudeness" in the current O(1) scheduler code in 2.4-ac
that could result in poor latency under certain workloads.
The patch should be in a near future 2.4-ac although I will need to
update the preempt-kernel patch to take advantage of it.
Robert Love