[Linux-ia64] Itanium gets supercomputing software

Alan Scheinine scheinin at crs4.it
Thu Apr 17 03:56:48 EDT 2003


Greg Lindahl wrote:


Good compilers have instruction scheduling which do this on other
chips. While it's easier to understand what's going on when the
parallelism is explicit, you'll find that scientific codes get a
pretty amazing number of instructions per cycle on quite a few cpus
and compilers.

The promise of EPIC was that it would be eaiser to do this. You'll
have to talk to some compiler people to find out if they think it was
easier. The ones I know hate EPIC was a passion.

 =================================================

   I do not think there was a promise that getting efficiency would
be easier with EPIC.  My understanding of the situation is that
the logic of dynamic allocation of resources, that is, the various
tricks done in silicon, could not scale to a large number of
processing units on a chip.  That is, the complexity grows faster
than linear, much faster.  If you take that as a postulate, then
it is logical to conclude that optimization must move to the
compiler.  The problem is that writing a compiler to maximize
efficiency is difficult.  Fifteen years ago I heard a talk in which
it was claimed that compiler advances developed at universities
arrive in commercial compilers after a delay of ten years.  More
recently people tell me that the development cycle is shorter, but
nonetheless, writing optimizing compilers is a very difficult task.

Greg Lindahl wrote that "The ones [compiler people] I know hate EPIC
with a passion".  Why?  Do they say that the concept is wrong or
is that problem that they cannot meet their deadlines because of
the quantity of analysis that has been moved from the silicon to
the compiler writer?  This is not a rhetorical question, it would
be interesting to learn more details from the "compiler people".

By the way, it may be a good idea to develop more packages like
Atlas and FFTW which optimize themselves based on the actual computer,
since memory latency and other factors are variable.  But then,
optimizing through experimentation takes a long time.

 -- Alan Scheinine
_______________________________________________
Beowulf mailing list, Beowulf at beowulf.org
To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf



More information about the Beowulf mailing list