[Beowulf] how can I calculate peak performance of a cluster

Mark Hahn hahn at physics.mcmaster.ca
Tue Sep 27 12:46:16 PDT 2005

> One should be very cautious with stating/using/selling peak performance. 

the problem is that it's an simple, understandable number.

> This is a bit like saying that my 15-year-old Honda automobile has a 'peak 
> speed' of 200 miles per hour.  This is a true statement: if I drop my car 
> off a very high cliff, it will reach 200 miles per hour speed just before 
> hitting the ground (:-).

I get your point, but it's more like saying that your Honda could hit 200
if all friction in your engine, drivetrain and chassis were to disappear.
(presumably leaving you with traction somehow...)

peak/theoretical rates are mainly of use when evaluating efficiency,
as is the case here (top500).  I don't think there's anything wrong with

on a factual note, I find that I can hit about .95 flops/cycle on an opteron
with the obvious/naive code for daxpy.  (that's incache, of course, before 
memory bandwidth takes over.)  I was half-expecting to be able to launch
an sse mul and an sse add each cycle (alas, addpd and mulpd both have a 
throughput of 1 every 2 cycles...  room for improvement there eh?)

More information about the Beowulf mailing list