From: Randy Crawford Newsgroups: comp.parallel Subject: Re: The Future of Massively Parallel Machines, SIMD or MIMD approach? Date: 20 Dec 1998 06:46:57 GMT Organization: ISPNews http://ispnews.com Approved: bigrigg@cs.cmu.edu Message-Id: <75i6h1$hjk$1@encore.ece.cmu.edu> Originator: bigrigg@ece.cmu.edu hsh2@cov.ac.uk wrote: > I'm interested in the trend which favour MIMD over SIMD or vice versa, > and which approach will likely be used in the year 2000 and beyond. Almost no SIMD machines exist today. Therefore MIMD will be a much bigger influence in the decade to come. > I have looked at a few papers on the MIMD vs. SIMD debate but all seem > to move to the economic/ cost side of things. That's because it's an old debate that was pretty much over by 1992. SIMD lost. > If cost was NOT a factor, which is better? which has or could have the > better utilisation of memory and processors, communication, memory > coherance and synchronisation... etc.? I'm interested in the > architectural benefits rather then the finanical benefits of each approach. It's harder to map the average parallel solution onto a SIMD architecture or language. MIMD is more usable in general, and therefore is more productive to most users. The advantages of SIMD in those rare ideal occasions just wasn't enough to keep it alive. > I know MIMD and SIMD are broad classifications, but would anyone like to > comment? Examples would be useful. You might look into the SPMD model (Single Procedure Multiple Data), where you have each processor run the same procedure at the same time, generally splitting up the data into equal partitions and passing each to a different CPU. IMO, SPMD has completely replaced SIMD, and may become a bigger influence in years to come if SPMD-related hardware/languages emerge soon (and there are indications that this may happen (the growth of POSIX threads-based parallelism; the potential of SPMD-like localized processing architectural models like Patterson's IRAM). SPMD is fairly easy to program, easier than MIMD to debug, and basically as efficient as SIMD while more generally applicable. But SPMD is really just a constrained variant of MIMD, so you'll have to decide whether SPMD deserves any ballyhoo in its own right. -- Randy Crawford crawford@net66.com N=1 ==> P=NP -- Articles to bigrigg+parallel@cs.cmu.edu (Admin: bigrigg@cs.cmu.edu) Archive: http://www.hensa.ac.uk/parallel/internet/usenet/comp.parallel