Re: First bench of 2.4-HEAD branch...

Henrik Nordstrom wrote:
>
> Joe Cooper wrote:
>
> > I've finished up the four runs I planned for this configuration of
> > HEAD. Out of curiosity I'm doing one more at an even lower req rate,
> > after that I'll start on other configurations in a couple of hours. The
> > first four runs are posted on my site:
>
> Thanks. It proved the current tuning of aufs is really weak and needs a
> serious boost to be effective.
>
> Here is something for you to try:
>
> change MAGIC1 in src/fs/aufs/store_asyncufs.h to read
>
> #define MAGIC1 NUMTHREADS*100
>
> This should make aufs a lot more aggressive in using the disk.

Will give it a go this afternoon. May be of interest, or not, but
upping the threads to 48 did make some difference for the better on
performance, as expected, but still left it lagging behind 2.2S5
significantly.

> It would also be nice if you could do a benchmark of hno-asyncio from
> sourceforge. The hno-asyncio branch includes a different I/O sheduler
> which should make a lot higher utilization of the threads while
> hopefully wasting less CPU time.

Per your other mail, I assume you'd like me to wait a day or two on this
one? If not, I'll fetch it tonight and run it on a second machine.

> Another related request: If would be very nice if you included some
> basic systems monitoring in your benches. It tells us a lot how far the
> system is loaded and in what ways. This is especially important when
> comparing two different implementations like the current 2.4-HEAD and
> hno-asyncio.

I was doing this a while back but didn't set it up on the new machine.
Forgetfulness on my part. Will attach this output from now on to the
benchmarks. Good idea, thanks.

I'll update you in a day or two when I've got some more results.
                                 --
                    Joe Cooper <joe@swelltech.com>
                Affordable Web Caching Proxy Appliances
                       http://www.swelltech.com
Received on Mon Jun 12 2000 - 14:25:13 MDT