On Sep 14, 2006, at 12:14 PM, Gil Bloch wrote:
Hi Brent,
You do not mention what interconnect (protocol or advertised
throughput) that you are considering. If you are considering
gigabit Ethernet, you can skip the rest of this message.
If, however, you are considering something faster than gigabit
Ethernet, bear in mind that GPFS only can use TCP (as far as I
know). You will be limited to the performance of the TCP on your
interconnect. If you are using TCP/IP/Ethernet, then expect a very
high CPU load. If you are considering an IB solution, look closely
at their IPoIB results.
Luster File System (open source from CFS) support native
InfiniBand, delivering higher performance with low CPU utilization.
Gil Bloch
Mellanox Technologies
Hi Gil,
I had Lustre in mind when I wrote that, but I could not speak for the
other filesystems mentioned because I do not know about them. Also, I
did not want it to appear that I was recommending one over another.
<waving the corporate flag>
That being said I just submitted the code for Lustre over MX and the
results are quite good. See https://mail.clusterfs.com/wikis/lustre/
MX-10G. By the way, those performance numbers are using a Myrinet
switch, but I can get the same results using a low-latency Ethernet
switch like a Fujitsu XG700.
Starting the MX support for PVFS2...
</waving the corporate flag>
To be clear, my employer is agnostic about which filesystem people
use (i.e. Myricom does not prefer/recommend any particular solution)
as long as people buy our NICs (and switches when needed). :-)
Scott
_______________________________________________
Beowulf mailing list, [email protected]
To change your subscription (digest mode or unsubscribe) visit
http://www.beowulf.org/mailman/listinfo/beowulf
