NAS Parallel Benchmarks for Current Hardware

andy andy at
Mon Jun 16 04:44:32 EDT 2003

Egan Ford wrote:
> > Unfortunately, I am not sure linpack and spec reliably measure what is
> > important for numerical work on a beowulf.
> I disagree.

I admit I should have been more careful and inserted the word "our"
befor "numerical work on beowulf". My concern is partly that they does
not test what I want to know and partly that linpack and spec blob
ratings shift a lot of hardware. 

What I need (and I supect many others but I will not speak for them this
time) is:

* a range of grid sizes so that I can see cache/memory effects 

* a range of common numercal solvers (covering most of those I use) with
differing communication characteristics so that I can gauge the likely
behaviour of a range of simulations.

No single number can provide such information but something the like NAS
parallel benchmarks do. 

15-20 years ago linpack stopped being a measure of anything useful when
manufacturers adapted their compilers/hardware to generate
unrepresentatively large blob ratings. This was brought home when
commissioning a small shared memory machine and we could not get near
the quoted linpack blob rating using various combinations of
optimisation flags (just over 10 blobs instead of nearly 30 blobs if I
recall correctly). After a phone call we used the undocumented -linpack
compiler flag and bingo. 

Evidence is growing that the spec blob ratings have suffered in a
similar way (though to a lesser degree). See, for example, "Sun breaks
another SPEC benchmark" in comp.benchmarks:
Beowulf mailing list, Beowulf at
To change your subscription (digest mode or unsubscribe) visit

More information about the Beowulf mailing list