James.Price.Salsman@cat.cmu.edu (08/07/89)
In article <1989Aug4.180349.3480@esegue.uucp> deb@quimby.dartmouth.edu (Deb Banerjee) writes: > Your suspicion about software rather than hardware being the challenge in > parallel programming systems may be correct. > > I am working on high level programming languages & systems for SIMD machines > (I have the Connection Machine in mind). > > References: >... Also, don't forget Guy Blelloch's master's thesis, "AFL-1: A Programming Language for Massivly Parallel Computers," MIT-AR-914(?). He defines a simple "Activation Flow Language" and builds a production system on top of it. Fast, fine-grained, totally-parallel production systems work well on SIMD systems, but the major implementation roadblock is the construction of a compiler that performs the spatial reasoning task of fitting all the flow channels together and mapping out the productions onto physical PE's. Also, bit-wide flow channels (as in AFL-1) are a bad idea for serious production systems. The "activation stream" should be wide enough so that the production's binding information can be sent through. Modern production systems like OPS5 have complex binding semantics, and most software written on top of them are dependent on those semantics. :James Disclaimer: The University doesn't think that I know what I'm doing. -- :James P. Salsman (jps@CAT.CMU.EDU)