You have a bottleneck by using only one metadata server. Make all 32 nodes BOTH i/o and meta data servers. Grouping the servers like you did is a GOOD thing.
Becky -- Becky Ligon PVFS Developer Clemson University 864-650-4065 > HI, all > > Now I have 32 nodes, both are used as IO servers and clients, one as > metadata server. When I write a different > > file in each client(e.g. client1 write /mnt/pvfs2/file1, client2 write > /mnt/pvfs2/files), the aggregate io throughout tested by > iozone(read/write) > is only half of performance of 4 nodes(both as IO servers and clients). > > I think each file are striped to 32 nodes will cause the additional > network > and data transaction overhead. > > If each file are only striped to small scale nodes(such as only 4 nodes > among 32), the performance will be improved ? > > But when I use twod_stripe mode, which divides the 32 IO servers into 7 > groups, each group strip factor is 256 and stripe size is 32M, the > performance is not improved as expected. > > Would you please tell me how to control each file to be stored into > different IO servers or improve the pvfs2 performance in 32 nodes by > tuning > parameters or others? > > _______________________________________________ > Pvfs2-users mailing list > [email protected] > http://www.beowulf-underground.org/mailman/listinfo/pvfs2-users > _______________________________________________ Pvfs2-users mailing list [email protected] http://www.beowulf-underground.org/mailman/listinfo/pvfs2-users
