Messages in this thread | ![/](/images/icornerl.gif) | | Date | Wed, 31 Dec 2003 22:03:54 +0100 | From | Roger Luethi <> | Subject | Re: 2.6.0 performance problems |
| |
On Wed, 31 Dec 2003 12:21:19 +0100, Jens Axboe wrote: > > Thanks. 2.5.39 alone will do, actually. I'm just curious how far the > > similarity between qsbench and bk export goes. > > 2.5.39 is when the deadline io scheduler was merged. How do you define > the qsbench suckiness?
2.5.39 was the biggest regression for qsbench (fixed later, most notably in 2.5.41). 2.5.39 was a significant improvement for efax ("fixed" in 2.5.43).
All I'm doing here is reading the graph I posted at: http://hellgate.ch/bench/thrash.tar.gz
For the systematic testing, I used "qsbench -p 4 -m 96" on a 256 MB machine. This allowed the kernel to achieve high performance with unfairness -- that's what 2.4 does: One process dominates all others and finishes very early, taking away most of the memory pressure. The spike for qsbench in 2.5.39 remains if only one process is forked (-p1 -m 384), though.
I asked for the bk export numbers with 2.5.39 because I'm curious how close to qsbench the behavior really is.
> Do you have numbers for 2.4.x and 2.6.1-rc with > the various io schedulers (it would be interesting to see stock, > elevator=deadline, and elevator=noop).
I planned to compare the io schedulers in 2.6.0 anyway. Do you expect different numbers for a recent bk snapshot?
Roger - 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/
| ![\](/images/icornerr.gif) |