BS"D
To those hardware oriented:
We have a compute cluster with 23 nodes (dual socket, dual core
Intel servers). Users run simulation jobs on the nodes from the head
node. At the end of each simulation, a result file is compressed to
2GB, and copied to the file server for the cluster (not the head
node) via NFS. Each node is connected via a Gigabit line to a
switch. The file server has a 4-link aggregated Ethernet trunk (4Gb/
S) to the switch. The file server also has two sockets, with Dual
Core Xeon 2.1GHz CPU's and 4 GB of memory, running RH4. There are
two raid arrays (RAID 5), each consisting of 8x500GB SATA II WD
server drives, with one file system on each. The raid cards are AMCC
3WARE 9550 and 9650SE (PCI-Express) with 256 MB of cache memory .
When several (~10) jobs finish at once, and the nodes start copying
the compressed file to the file server, the load on the file server
gets very high (~10), and the users whose home directory are on the
file server cannot work at their stations. Using nmon to locate the
bottleneck, it appears that disk I/O is the problem. But the numbers
being reported are a bit strange. It reports a throughput of only
about 50MB/s, and claims the "disk" is 100% busy. These raid cards
should give throughput in the several hundred MB/s range, especially
the 9650 which is rated at 600MB/s RAID 6 write (and we have RAID 5).
1) Is there a more friendly system load monitoring tool we can use?
2) The users may be able to stagger the output schedule of their
jobs, but based on the numbers, we get the feeling the RAID arrays
are not performing as they should. Any suggestions?
Thanks
Harry
------------------------------------------------------------------------
-
Harry M. Greenblatt
Staff Scientist
Dept of Structural Biology [EMAIL PROTECTED]
Weizmann Institute of Science Phone: 972-8-934-3625
Rehovot, 76100 Facsimile: 972-8-934-4159
Israel
