[Gluster-devel] Write speed problem in single client - single server setup

[Dmitry Unkovsky]

Hello.

We are trying to setup a redundant glusterfs storage. But the main
problem in our case is a write speed, even with single client - single
server setup. With read speed about 75-80MB/sec, write is at best
~30MB/s
from a single client (even with a server located on a local node
trough a unix/socket) with
relatively big files. Background writing on a remote node (which
happens with afr, for example) gives an additional slowdown. In a real
life clustered environment with concurrent access things can be a
little bit different, but the results we experience now do not look
good, anyway.

We've tried nearly any combination of performance translators
options, booster, parallel and sequential writing with /bin/cp, /bin/dd
with different block size and such.

A quick, not too deep look at the code was not enough to find a place
that can cause such drop of write performance comparing to read. Can
you give a
hint where to dig further or what could be tweaked in the code to
improve this situation? This could be our simple mistake or
or some misunderstanding.

Also, it would be nice if someone would share his results on read/write
speed in a simple one client / one server setup (tcp or unix socket)
with and without performance translators. This wold give an idea of
what others can expect and where is a point to stop tunning things =)

Test systems have current version from tla with fuse-2.7.2 patched with
a little bit adjusted patch for fuse 2.7.0. They are AMD Athlon64
4200+/2GB RAM/200GB SATA 7200RPM HDD each, connected with gigabit
Ethernet. Host itself is capable of sequential reading from HDD at
~100MB/s, writing at ~90MB/s. Ftp reads are around ~90MB/s, writes
~75MB/s from a remote host.

P.S.
Thank you for your project!