Linux 2.6.23 Kernel Benchmarks

I've been running 2.6.23 since it came out, and it's been very smooth. The only compile option I changed was to set the HZ to 1024 rather than the default 250, and some tweaks in sysctl.conf:

kernel.sched_min_granularity_ns = 1000000 kernel.sched_latency_ns = 10000000 kernel.sched_compat_yield = 1 kernel.sched_wakeup_granularity_ns = 1000000

But I am always looking down the road, so I downloaded the -rt patcheset for 2.6.23 and started the compile.

I've got azureus running with 6 torrents seeded, and since the box is busy with bt and a kernel compile, I decided to get online with a game of quake 3 arena to see how she handled under load.

While it was not quite as smooth as usual, the game was quite playable (and in fact I won handily) While this sort of thing wouldn't show up in a typical benchmark, I'm nonetheless impressed by the improved feel.

So 'responsiveness' is good?

> So 'responsiveness' is good?

Indeed so. I haven't actually gotten around to booting into the -rt kernel, since I'm stlil having so much fun with the stock 2.6.23.

Not too long ago, the idea of playing a 3D FPS internet game while compiling the kernel would have been a silly and naive idea - or a game for masochists - let alone hosting 6 azureus torrents at the same time.

herzeleid -

Thanks VERY much for sharing your tweaks. I haven't had a chance to try them myself, but am hoping they will smooth out my myth operations.

I run a little close to the ragged edge running HDTV over 80211G, and it isn't helped by mythbackend's ginormous sucking memory leak.

Me hopes that, until that sucker is fixed, some tweaks will make paging a and task switching a little less disruptive to my backend.

Dino, did you read this?

http://rudd-o.com/archives/2007/10/02/tales-from-responsiven...

It may well be this helps, because myth uses large files. Also, be sure to use XFS or JFS as your filesystem (I suppose you did that).

Hans -

Already set both of those, but it ain't enough on my backend box, which must share duty with other uses.

I do use JFS for the myth stuff.

> I do use JFS for the myth stuff.

Is it mounted with the -noatime option?

BTW the testing I did on SLES 10 showed jfs as being pretty easy on the CPU but its performance was fairly abysmal. xfs might be a better choice IMHO.

>Is it mounted with the -noatime option?

I don't think so, but I also don't think it's my problem. I have a frontend/backend setup. The frontend is dedicated, but the backend is my regular workstation. The problem, I think, is related to a memory leak in the backend that causes it to grab great quantities of memories, which, in turn, generates paging.

At least that's my guess. Things tend to get better when I stop/restart the mythbackend.

My hope is that, until the memory leak gets resolved, that we can smooth out overall operations enough to notice.