[Beowulf] benchmarking network and performance

Robert G. Brown rgb at phy.duke.edu
Thu Jan 15 08:48:49 EST 2004


On Wed, 14 Jan 2004, Jim Lux wrote:

> Thanks.. I suspect that measurements vs packet size are of some interest.
> (something like HINT, except for size instead of computation).

netpipe.  netperf, too, but nobody loves it any more.  netpipe will let
you transparently test raw TCP, PVM and MPI connections side by side,
for varying packet sizes, with streams of packets long enough to get
decent statistics on them, and autogenerate a nice little report on
same.

Most of what you learn from such a test on ethernet is that small
packets are latency bound and the interface sends as many of them per
second as it can (given the latency) largely independent of packet size.
Consequently bandwidth grows linearly with packet size.  At some point
the volume of data being transmitted STARTS to actually run into the
wire bandwidth (if you like, the linear growth curve gets close to the
wirespeed saturation/constant line) and the linear growth curve gently
bends over and saturates sort-of-exponentially, typically at some 90% of
wirespeed (for a decent interface), typically when packet sizes reach
roughly 1K in size (maybe 2/3 of the standard MTU).

Sometimes there is interesting "structure" in the curve -- "often" and
"egregious" structure for poor interfaces, less often for good ones but
sometimes you'll see something like dips or peaks around certain packet
size boundaries.

Oh, and lmbench has a variety of network latency and bw tests (tcp and
udp and localhost) mixed in with the rest of its microbenchmarks.  Maybe
a bit clunky to use compared to netpipe but you wanted other benchmark
tools as well and this is a suite of them.  Finally, for other benchmark
tools there is cpu_rate, which includes stream (with malloc and a sweep
of vector sizes) as well as the savage benchmark and some other
stream-like arithmetic tests.  If your organization participates in SPEC,
there is SPEC, and of course the list archives contain pointers to
various parallel performance benchmarks.

   rgb

> 
> _______________________________________________
> Beowulf mailing list, Beowulf at beowulf.org
> To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf
> 

-- 
Robert G. Brown	                       http://www.phy.duke.edu/~rgb/
Duke University Dept. of Physics, Box 90305
Durham, N.C. 27708-0305
Phone: 1-919-660-2567  Fax: 919-660-2525     email:rgb at phy.duke.edu



_______________________________________________
Beowulf mailing list, Beowulf at beowulf.org
To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf



More information about the Beowulf mailing list