[PRL] POPL Trip Report

William D Clinger will at ccs.neu.edu
Thu Jan 10 09:32:25 EST 2008


Concerning CUDA, Mitch wrote:
> Sounds like neat stuff.  Do we have anybody interested in this?

Gene Cooperman.  He had some role in organizing a workshop
on this at NU a few months back.  I attended that workshop
in the morning, but got bored and left after lunch.

> Anybody who has large parallelizable computations for which this might
> be useful?

The GPU cards are designed for massively data-parallel
floating-point algorithms.  The newer cards are versatile
enough for novel uses, e.g. a snapshot-at-the-beginning
algorithm that marks all live objects in a heap by having
every word of memory check to see if it is a reference
that has been marked; if so, then set the mark word for
the word it points to; iterate that process to a fixed
point.  The problems are that the cards don't have enough
memory for algorithms like the one I just described, and
the cost of copying large amounts of data between memory
systems is hard to overcome.

Will



More information about the PRL mailing list