Date: Sun, 6 May 2007 02:10:53 -0400
From: Ron <despair@adelphia.net>
To: linux-kernel@vger.kernel.org
Cc: ck@vds.kolivas.org, mingo@elte.hu, despair@adelphia.net
Subject: CFS v9 vs SD 0.48
Message-ID: <20070506061053.GA16399@comcast.net>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
User-Agent: Mutt/1.5.15+20070412 (2007-04-11)
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 4153
Lines: 75

Here's my experiences with CFS v9 and -ck1. Base for the two kernels is
2.6.21-git4-nohz (x86_64 nohz patches I snagged off LKML a month or so back).
The extra features in CK shouldn't matter, most of the tests are running on
an otherwise unloaded system, with the apps already warmed up.

Let's start with some game benchmarks.

Doom 3 is run at 640*480@60Hz with custom settings roughly near high quality.
QuakeForge is run at 1680x1050@60Hz with max quality settings and a customized
replacement texture set based on the QRP set.
( http://facelift.quakedev.com/retexture )
                doom 3 demo1            QF overkill
-cfs9  X -10    66.9 fps                181.7 fps +-0.040 fps
-cfs9  X  0     67.6 fps                181.5 fps +-0.025 fps
-ck1   X -10    67.1 fps                182.1 fps +-0.063 fps
-ck1   X  0     67.9 fps                182.5 fps +-0.094 fps
Average is calculated from 5 runs, throw out low & high, average remaining 3.
Exception: Doom 3 was EXTREMELY variable under CFS v9 at X -10, and had nasty
lower average framerates as low as 64 fps on a few runs. So Doom 3 for X at
-10 was calculated by throwing out the high & low of 9 runs, and averaging
remaining 7. It's nicely consistent under SD regardless of X renicing.
Unfortunately I forgot to record the deviation for Doom 3, but it was around
0.08fps, except for CFS v9 with X reniced runs. Note the variance is without
outliers, with outliers it looks much worse for cfs.

Both are quite resistant to music stalls or sputtering, with a music player
going during the compile.

It seems to me that renicing is dangerous under CFS. It can easily cause
horrendous latencies. With the compile mentioned above going in one
gnome-terminal, and waving the pointer around (sloppy focus/follows pointer)
and typing, I can easily wind up with text not appearing for 4-5 seconds, or
even losing characters, under CFS with X at -10. Everything else is running at
default priority. Unreniced, CFS is better than mainline, but still has small
lags in mouse movement during the compile & wave test.

Under SD, which has less of a delay to window borders being redrawn, it gets
even better with X at -10. Other than that there's no overt penalty for
varying nice. Latency gains for renicing X aren't as large as with older SD.

The harshest test for CFS is games running in wine. Diablo II: Lord of
Destruction in winehq 0.9.36 doesn't stall audio like older versions of CFS
did... Instead it stalls rendering during heavy combat. As unpleasant as
static and dropouts are to listen to, character death due to badly behaved
scheduler is worse.

Guild Wars under Cedega (unplayable in winehq on my machine, or I'd use that)
CFS gets horrible stalls during heavy action, and also unpredictably during
normal play. Also prone to audio breakup.

SD runs D2 & GW perfectly smoothly, and without any audio glitches at all. I've
had zero audio glitches in 2.6.21-ck1 even when deliberately trying to provoke
them. I haven't resorted to negatively niced parallel make, yet.

My system is a socket 754 Athlon 64 3000+, 1GB PC3200, NForce3 250 motherboard,
GF6800GT using Nvidia's 100.14.03 beta drivers. I'm using libata with CFQ disk
scheduler for my ATA133 & ATA100 hard drives. Ubuntu x86_64 Gutsy Gibbon with
a 32bit chroot for old games and all the web cruft that doesn't have 64bit
support yet. USB trackball is polled at 200Hz.

On a 10 scale, I'd rank SD a 9, windows around 6, CFS 5, and the stock
scheduler 3 (up from a 1 last year), for interactivity and load handling. I've
used linux by preference for years, but UI latency has not been one of it's
strong points. Nor resistance to audio underruns on my systems. Until now.
Bottom line, SD won on everything I tested. Which were all picked from things
stock used to trip over. CFS doesn't entirely fix them, it shifts where and how
badly.

Thanks!
Ragnvald "Despair" Maartmann-Moe IV
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/