Could you compare the following three ways to copy a file onto PVFS
(without ftp or samba for the moment) to see if it is specific to "cp"?
a) time cp /tmp/file.dat /mnt/pvfs/file1.dat
b) time dd if=/tmp/file.dat of=/mnt/pvfs/file2.dat bs=8M
c) time pvfs2-cp /tmp/file.dat /mnt/pvfs/file3.dat
If there is a big difference between a) and b), it may just be that your
version of cp (ie, coreutils package) is using a 4KB block size
regardless of what the file system reports. If there is a big
difference between b) and c), it may indicate a kernel module problem.
If all three are slow then there must be a more fundamental issue.
-Phil
Emmanuel Florac wrote:
Hi, I've made a pvfs2 cluster to agregate storage from 2 machines. The
machines run Linux Debian etch with 2.6.22.19 linux. All is well, both
machines are identical with 3Ware raid controllers, and write disk
thruput is about 550MB/s.
Each machine has 4 GigEth ports. I've dedicated 2 of them to link the
machines together (bonding, 802.3ad) for the pvfs traffic, the 2 others
are connected to the outside world. The purpose is to provide a rough
ftp/samba storage thru the VFS layer, and it works fine.
However the performance is abysmal : the network thruput between the
nodes is 8MB/s while copying! the total cluster write performance
(cp from local disk to pvfs storage) is 17 MB/s.... What can be going so
wrong? Any hint would be appreciated.
_______________________________________________
Pvfs2-users mailing list
[email protected]
http://www.beowulf-underground.org/mailman/listinfo/pvfs2-users
