eugene@ames-lm.UUCP (Eugene Miya) (05/22/84)
Ask the man (or woman) that owns one!
We have the 3rd XMP. I have just started programming on it. First,
comparisons between the XMP and VP are somewhat bogus. Don't forget
the the basic technology of the XMP and its software are 1975.
The CFT compiler only vectorizes inner DO loops. We have an inhouse
language called vectrall which vectorizes two levels of loops (I haven't
used it yet).
Kenji Muira (U Ill 1970s) stopped by to describe the VP-100/200 line.
It is not the FORTRAN compiler which is good about the VP, it is the
interactive optimizer which suggests optimizations. It is not
perfect, just a cut above what Cray has. Cray is working on a new
Fortran compiler (NFT) for certain computers running certain OSes
which can't be described in this forum, but nothing like this optimizer.
It appears the optimizer is based on the work pioneered by David Kuck
and others in the early-mid 1970s (called Parafrase).
Other comments: APL, this has been suggested by others here: problems:
existing dusty deck FORTRAN programs, the brains of those physicists
who learned FORTRAN either on their own or in Physics Dept. FORTRAN
classes. APL represents another cryptic tool (worse than UNIX they think)
to learn. Some are willing, but remember they are here to do physics, not be
students of programming languages.
Vectorization: The problem is not just using a vector language. Many
parts are very serial. The Cray succeeded because it was a very fast serial
machine as well. The problem is inherent in any of the algorithms
(you know those scalar constants, etc.). Lastly, you have to deal with
those long execution times (how about months of Cray time?).
Our people don't want billions of floating point operations (note: no time
measurement), they want trillions and quadrillions of operations!
Remember the machine "Deep Thought" in Hitchhiker's Guide to the Galaxy,
that gives a favor of the realm of some of the problems run on Crays.
Last word: if you want more, I think I will be leading a Bird of the Feather
session at Usenix in June.
--eugene miya
NASA Ames Research Center
{hplabs,hao,research,dual}!ames-lm!statvax!eugene