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.
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.
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.
A simple "vmstat 1" log is sufficient for my needs.
while true; do
date
uptime
vmstat 1 600
done >vmstat.log
You also need to note the start and stop time of the polygraph run so
the polygraph graphs can be correlated to the vmstat data.
/Henrik
Received on Mon Jun 12 2000 - 05:59:17 MDT
This archive was generated by hypermail pre-2.1.9 : Tue Dec 09 2003 - 16:12:29 MST