FastE NIC Recommendation (latency)

[Robert G. Brown]

On Mon, 18 Nov 2002, Michael Stein wrote:

> > Unlike the case of a global shared memory, low latency is not always a
> > good thing.  Gigabit Ethernet adapters go to some effort to _increase_
> > latency so that the CPU can process multiple messages during each
> > interrupt.  This interrupt mitigation might even be clever enough to
> > look at the destination of the next incoming packet to decide if the
> > interrupt should be deferred.
> 
> That must be what I'm seeing as I was just using "ping".
> 
> What's a good way to measure latency?

There are a number of decent tools for measuring both bandwith and
latency.  There is a suite of microbenchmarks you might look into
(lmbench) at:

   http://www.bitmover.com/lmbench/lmbench.html

which is distinguished by being used by Linus Torvalds (among many
others:-) for microbenchmarking kernel features in the development
series.  It contains a variety of network bandwidth and latency
microbench tools, as well as memory latency/bandwidth, disk l/b, and
much more.

Another one that is often used is netpipes:

  http://www.scl.ameslab.gov/netpipe/

which has the advantage of being protocol independent -- the same suite
measures latency/bandwidth in TCP, MPI, PVM contexts (for example).

An older one that is no longer being maintained (I don't think, correct
me if I'm wrong) is netperf:

  http://www.netperf.org/netperf/NetperfPage.html

The site still exists, but last time I checked the source hadn't changed
in years and wouldn't compile without significant hacking on modern
libc6 linux boxen.  I believe I did the hacking required and that it
does still work, but netpipes or lmbench both are actively being further
developed and improved and hence are likely better choices.

   rgb

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