Hi,
I got the latest pvfs from CVS and rebuilt both the server and the client with the "--without-bmi-tcp" flag. This took me from 260MB/sec to about 320MB/sec with pvfs2-cp. About a 60MB/sec boost. adjusting the buffer size option to pvfs2-cp did help: pvfs2-cp -b16777216 -t /tmp/testfile1 /test/pvfs2/foo12 Wrote 16777216000 bytes in 45.138288 seconds. 354.466257 MB/seconds pvfs2-cp -b33554432 -t /tmp/testfile1 /test/pvfs2/foo13 Wrote 16777216000 bytes in 43.276829 seconds. 369.712855 MB/seconds pvfs2-cp -b67108864 -t /tmp/testfile1 /test/pvfs2/foo14 Wrote 16777216000 bytes in 40.825193 seconds. 391.914866 MB/seconds pvfs2-cp -b104857600 -t /tmp/testfile1 /test/pvfs2/foo17 Wrote 16777216000 bytes in 39.127303 seconds. 408.921615 MB/seconds pvfs2-cp -b134217728 -t /tmp/testfile1 /test/pvfs2/foo15 Wrote 16777216000 bytes in 41.528605 seconds. 385.276606 MB/seconds Best I can get is around 400MB/s for a one(client) to one (server) test config. This is really close to what my disk back end can handle. Since there is only one I/O server using the -s option for striping did not help any in this case. Thanks Rene On Thu, 2007-09-06 at 19:16 +0100, Scott Atchley wrote: > Hi Rene, > > First, what is your FlowBufferSize set to in your fs.conf file? > > Second, I found that by playing with pvfs2-cp's switches, I could > improve performance some. I typically use a FlowBufferSize of 4 MB. > Pete would have to comment on what he uses with IB. Given that, I set > -b $((N*1024*1024)) where N is 16, 20, 32, 40, 64, 100, and 128. You > may also want to try using -s $((M*$FlowBufferSize)) where M is 1, 2, > 4, etc. > > Scott > _______________________________________________ Pvfs2-users mailing list [email protected] http://www.beowulf-underground.org/mailman/listinfo/pvfs2-users
