[Beowulf] 10G networking?
landman at scalableinformatics.com
Fri Jan 23 09:41:36 PST 2015
On 01/23/2015 12:14 PM, Gavin W. Burris wrote:
> On 12:04PM Fri 01/23/15 -0500, Joe Landman wrote:
>> Don't use ping for HPC latency/throughput tests.
> What is your preferred benchmark, Joe?
Ping isn't a benchmark. Its a very basic test of connectivity, packet
transfer and routing. HPC folks generally use it for baseline sanity
checks, not performance.
HPC folks would likely use an MPI benchmark similar to
http://mvapich.cse.ohio-state.edu/benchmarks/ for application stack
performance similar to what Mark was talking about ("min packet
ping-pong latency is the standard metric, though packets/sec
rate might be interesting as well").
Mark is dead on correct pointing out that ping-pong latency is one of
the standard metrics. It is "roughly" correlated with latency sensitive
code performance, at least the sections that are latency sensitive.
You could use netperf or iperf for throughput tests on ethernet. I use
the latter to see the effects of scaling. I usually run it as
iperf -t 60 -P 4 ...
You can also use other similar tests, and there are a fairly wide
ranging number of them.
For IB and RoCE/iwarp you can use the rdma diagnostics, which depending
upon your distribution, could be named things like (regex)
root at usn-ramboot:~# ib_
ib_clock_test ib_rdma_lat ib_read_lat
ib_rdma_bw ib_read_bw ib_send_bw
[root at CentOS6-build ~]# rdma_
rdma_client rdma_server rdma_xclient rdma_xserver
The man pages and help are useful for these.
But, not Gavin, ping isn't normally reported when one talks about
ping-pong. See the OSU link for the actual test. Specifically,
osu_latency and osu_latency_mt for contention.
Joseph Landman, Ph.D
Founder and CEO
Scalable Informatics, Inc.
e: landman at scalableinformatics.com
p: +1 734 786 8423 x121
c: +1 734 612 4615
More information about the Beowulf