I/O Error & Compiler Question

Greg Lindahl glindahl at hpti.com
Thu Jan 4 22:05:31 PST 2001


>     Also, What compilers are available that will parallelize our
> existing code?  Right now the only reason that there isn't a large
> cluster, is that we can't find a compiler ( like other vendors ) that
> will parallelize our code w/ out re-writing it.
> Money isn't a concern w/ the compilers.

There is no magic bullet. I believe that the PGI compilers have an
autoparallelization option but it generally doesn't work so hot with most
codes. You might get a 2x or 3x speedup, but only with datasets so huge that
the problem has to be essentially embarrassingly parallel.

Since you're from NOAA and are probably doing weather, it's worth noting
that there is a system called SMS developed at the Forecast Systems Lab in
Boulder. It only works on stencil codes (like weather forecast codes), and
allows you to add a relatively small number of compiler directives to your
F77 code, and it does the rest of the work. The nice thing is that it gets
speedups as good as 100x on 100 cpus for real codes on real datasets.

http://www-ad.fsl.noaa.gov/ac/sms.html

[ I'd also love to see some non-weather people with stencil codes, which are
extremely common for finite difference solutions, try SMS... ]

-- greg





More information about the Beowulf mailing list