[Beowulf] curiosity killed the cat
lindahl at pbm.com
Thu Jun 2 09:36:17 PDT 2016
Yes, I'm pleased that several companies have adopted message rate as a
metric, although I'm also sad that some have gamed the benchmark with
N1/2 was intended to say something about the shape of the
bandwidth-by-message-size curve. The smaller it is, the farther
you can go with strong scaling.
Historically I had lobbied for people to publish N1/2 Linpack numbers,
never got anywhere with that. N1/2 and message rate were my attempt to
update the usual "bandwidth and latency" microbenchmarks to something
with more meaning.
On Wed, Jun 01, 2016 at 03:29:51PM -0700, Holger Fröning wrote:
> Greg, we all do remember n/2 (actually I still teach about it). If I remember right, PathScale introduced the message rate metric, too, and that actually became vey popular. Personally, I prefer this one as it is very descriptive. Personally, I found it difficult to compare different technologies using n/2.
> > On 01 Jun 2016, at 15:18, Greg Lindahl <lindahl at pbm.com> wrote:
> > On Wed, Jun 01, 2016 at 03:25:22PM -0400, Andrew Piskorski wrote:
> >> You mean GAMMA (Genoa Active Message MAchine):
> >> http://www.disi.unige.it/project/gamma/
> >> http://www.linux-mag.com/id/7253/
> >> http://www.beowulf.org/pipermail/beowulf/2006-February/014846.html
> > I wonder how their flow control scheme of static partitioning of the
> > receive queue scales to more modern hardware?
> > Oh, and I'm thrilled that they show the N_1/2 packet size! (the size
> > where bandwidth reaches 1/2 of the peak.) That's a number PathScale
> > published, but it never became popular with other folks doing
> > performance testing.
> > -- greg
> > _______________________________________________
> > Beowulf mailing list, Beowulf at beowulf.org sponsored by Penguin Computing
> > To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf
More information about the Beowulf