ian@argosy.UUCP (Ian L. Kaplan) (10/18/89)
In article <20336@princeton.Princeton.EDU> mg@notecnirp.edu (Michael Golan) writes: >Last year, I took a graduate level course in parallel computing here at >Princeton. [...] > >1) There is no parallel machine currently the works faster than non-parallel >machines for the same price. The "fastest" machines are also non-parallel - >these are vector processors. > Clearly this is proof that a little knowledge is a dangerous thing. Even n-cube machines run applications like Monte Carlo simulation with _much_ better price performance than supercomputers. Now it might be claimed that this is a special class of applications. However parallel processors are not limited to n-cubes. The Connection Machine has beaten Cray machines on a number of classic vectorizable codes (e.g., fluid flow). For reference see "Proceedings of the Conference on Scientific Applications of the Connection Machine", 1988, Edited by H. D. Simon, World Scientific press. Note that the Connection Machine is probably less than half the cost of the Cray. I am sure that even cheaper SIMD processors will appear in the near future. >2) A lot of research is going on - and went on for over 10 years now. As far >as I know, no *really* scalable parallel architecture with shared >memory exists that will scale far above 10 processors (i.e. 100). And >it does not seems to me this will be possible in the near future. By this narrow deffinition, the statement is more or less correct. Classic shared memory MIMD systems with snoopy caches saturate rapidly. However, this is simply the wrong approach to the problem. SIMD architectures like the Connection Machine are scalable. Perhaps you did not study these. >3) personally I feel parallel computing has no real future as the single cpu >gets a 2-4 folds performance boost every few years, and parallel machines >constructions just can't keep up with that. It seems to me that for at least >the next 10 years, non-parallel machines will still give the best performance >and the best performance/cost. This statement is not even true for MIMD processors. Multiprocessor, with shared memory, used as file servers have _much_ better performance than comparable single processor systems. Then there are the issues of fault tolerance. When a Multiprocessor fails, you pull a board out and go on. When a uni-processor fails, you wait for field service. Sequent is doing a booming business selling shared memory multiprocessors. Even DEC is selling a multiprocessor (the VAXstation 3520). Ian Kaplan MasPar Computer Corp. argosy!ian@decwrl.dec.com The opinions expressed here are not necessarily shared by MasPar Computer Corp.
rang@cs.wisc.edu (Anton Rang) (10/18/89)
In article <308@argosy.UUCP> ian@argosy.UUCP (Ian L. Kaplan) writes: >In article <20336@princeton.Princeton.EDU> mg@notecnirp.edu (Michael Golan) writes: >>2) A lot of research is going on - and went on for over 10 years now. As far >>as I know, no *really* scalable parallel architecture with shared >>memory exists that will scale far above 10 processors (i.e. 100). And >>it does not seems to me this will be possible in the near future. > > By this narrow deffinition, the statement is more or less correct. >Classic shared memory MIMD systems with snoopy caches saturate >rapidly. However, this is simply the wrong approach to the problem. >SIMD architectures like the Connection Machine are scalable. Perhaps >you did not study these. The group working on the IEEE SCI (scalable coherent interconnect) claims that they will be able to handle up to 65,536 processors in either a message-passing or shared-memory environment. I haven't had a chance to read all their working papers, but it does seem that their stuff should scale well to 100-500 processors, at least. Their goal is 1 GB/sec bandwidth per processor (16 bits every 2 ns), and the interface chips are supposed to be available next year. (I'm eagerly waiting to see what this looks like... :-) The reference number for the standard is IEEE P1596.... Anton +----------------------------------+------------------+ | Anton Rang (grad student) | rang@cs.wisc.edu | | University of Wisconsin--Madison | | +----------------------------------+------------------+
tihor@acf4.NYU.EDU (Stephen Tihor) (10/19/89)
ACutally all of DEC's new systems above the desktop are multiprocessors. Gotta keep moving up to avoid them there killer micro's.
cik@l.cc.purdue.edu (Herman Rubin) (10/19/89)
In article <308@argosy.UUCP>, ian@argosy.UUCP (Ian L. Kaplan) writes: > In article <20336@princeton.Princeton.EDU> mg@notecnirp.edu (Michael Golan) writes: ....................... > Even n-cube machines run applications like Monte Carlo simulation > with _much_ better price performance than supercomputers. Now it > might be claimed that this is a special class of applications. > However parallel processors are not limited to n-cubes. The > Connection Machine has beaten Cray machines on a number of classic > vectorizable codes (e.g., fluid flow). For reference see "Proceedings > of the Conference on Scientific Applications of the Connection > Machine", 1988, Edited by H. D. Simon, World Scientific press. Note > that the Connection Machine is probably less than half the cost of the > Cray. I am sure that even cheaper SIMD processors will appear in the > near future. ...................... All computationally efficient means of generating non-uniform random numbers involve what are called acceptance-rejection or acceptance-replacement methods. These are most easily done on stream vector machines, and next on machines which have at least (vector-register)-memory transfer with non-rigid vectors, that is, moves in which the order of the items moved is fixed, but very definitely not the locations. Not all vector machines have this capability, and replacement is not vectorizable. The problem is worse with MIMD, although something is salvageable, but SIMD suffers from intrinsic problems. If the replacement procedure could be added to hardware, SIMD would only suffer a moderate penalty. -- Herman Rubin, Dept. of Statistics, Purdue Univ., West Lafayette IN47907 Phone: (317)494-6054 hrubin@l.cc.purdue.edu (Internet, bitnet, UUCP)
prc@erbe.se (Robert Claeson) (10/22/89)
In article <12780018@acf4.NYU.EDU> tihor@acf4.NYU.EDU (Stephen Tihor) writes: >ACutally all of DEC's new systems above the desktop are multiprocessors. Gotta >keep moving up to avoid them there killer micro's. A multiprocessor system is *not* the same thing as a parallel system. DEC's systems are symmetric multiprocessing systems (if now only Ultrix could run symmetrically multiprocessing) but *not* parallel systems. Don't confuse these two concepts, folks. Almost all parallel systems can function as symmetric multiprocessors as well. The inverse not always true. -- Robert Claeson E-mail: rclaeson@erbe.se ERBE DATA AB