[comp.parallel] Ultracomputer query

yoram@garfield.cs.columbia.edu (Yoram Eisenstadter) (02/13/89)

Can anybody tell me what kind of local memory has
been implemented on the N.Y.U. Ultracomputer?  My
latest reference, which is a bit dated, states
that the designers intended to place a large cache
on each processing element.  Is this what they've
actually done, or have they taken a different
approach (e.g., explicitly addressed local memory,
no local memory, etc.).

Thanks..Y
--
Yoram Eisenstadter                     | Internet: yoram@cs.columbia.edu
Columbia U., Dept. of Computer Science | UUCP:     rutgers!columbia!cs!yoram
500 West 120th Street, Room 450        | Bitnet:   yoram@cucsvm
New York, NY 10027                     | Phone:    (212) 854-8180

edler@cmcl2.NYU.EDU (Jan Edler) (02/20/89)

Yoram Eisenstadter (yoram@garfield.cs.columbia.edu) asks
about local memory in the NYU Ultracomputer.

The architecture has always included a cache for each processor.  In
addition, we have generally considered separately addressable local
memory to be a useful (but not crucial) extension to the architecture.
Indeed, the prototypes we've built to date include both cache and local
memory, but the local memory is fairly small (16KB) and used only for
diagnostics and a few rather specific operating system purposes.

If we had larger local memories, we'd begin to make it available to
users, possibly restricting its use to processes
that are non-preemptable, for simplicity.  We could move pages around
on demand as processes migrate, but our current feeling is that the
benefits of jumping through those hoops are uncertain.

User-accessable local memory would naturally be used for private
variables.  In addition, compilers could use it for temporary local
copies of shared variables.  Of course, these are also the purposes
served by the processor cache.  But there are cases, particularly in
large numerical computations, where software-managed local buffering
can be more effective than data cacheing given the current state of the
art in cache design (LRU).

Hope that helps,

Jan Edler
NYU Ultracomputer Research Laboratory
New York University
715 Broadway, 10th floor
New York, NY 10003
(212) 998-3353
edler@nyu.edu