Gluster users - May 2013 - Concurrent connection optimizations

Hi All,

I am experimenting with a replicated setup with 2 bricks per volume on
a shared webhosting implementation. Two nodes, each both a server and
client.

When doing some load tests, I see that the glusterfs process that is
the client mount seems to be the most CPU intensive of the various
things involved.

My tests have been somewhat inconclusive so far. I wonder if someone
could help make an educated guess as to what would offer better
concurrency performance.

Suppose I have one gluster volume "gv0" mounted on

/home/gv0/

and two virtual hosts www.vhost1.com and www.vhost2.com each being
served from this volume.

Suppose that I can concurrently serve 5,000 requests per second from
each of these virtual hosts, with the gluster client being the
bottleneck. (If I run bypassing gluster and read directly from my
bricks I can get 20,000). CPU runs about 120% at full-tilt according
to top. I have 16 cores to play with.

Would the concurrency be potentially higher if each of the virtual
hosts were on different gluster volumes? If so, and each had its own
client process, might I gain something by assigning specfic cpus via
cgroups or something?

Thanks,

Joseph