We are doing some research here in the brazilian government in distributed mass storage.
one of the scenarios that will be study involve gfs, and we will test gfs performance localy, exported to another machine via gnbd, iscsi and enbd.
we found some weird results, where my disks localy can reach over 300MB/sec using ext3, and the same test using gfs reach only 38MB/sec.
There is one graphic about this analysis in the page, and results for dbench and iozone.
In this moment we are doing benchmarks using bonnie++ to certify that this weird performance arent relate to the tools we had used.
Other consideration is that turning off hyperthread improved a litle (6-10MB/s) the performance, i believe this is a bug, or one problem in design of gfs.
is there a way to improve the gfs performance ? or this poor performance is all that i can get?