wailes@en.ecn.purdue.edu (Tom S Wailes) (06/23/91)
Assume you are designing a massively parallel computer system made of many commodity microcomputers. To make this interesting, assume that partitioning can be made to allow differing classes of users to coexist on one machine. For example, several "vi" tasks could be resident on a single processor, while many (128) processors could be attacking a electron transport simulation problem. What word size would be the best for use and why? A massively parallel machine would probably be used for scientific computations and many would say you must have a machine with 64 bit capabilties built in. However, all of the general purpose programming and low level processing tasks could easily be handled by present 32 bit micros, and clearly the 32-bit market will have the most competition and the most competitive price for some years to come. You must make a design decision soon. What papers have studied this? Who has numbers from which to base this decision? Tom wailes@ecn.purdue
davidsen@crdos1.crd.ge.COM (Wm E Davidsen Jr) (06/24/91)
In article <1991Jun23.012644.12449@en.ecn.purdue.edu> wailes@en.ecn.purdue.edu (Tom S Wailes) writes: | What word size would be the best for use | and why? Since the answer all depends on the definition of "best" I guess we're going to waste several weeks talking at cross purposes. Why not restate the question giving some hint what you're trying to maximize. You can go for max MIPS, MIPS/$ (with or without software cost included), lowest power drain, or best paint job. You can specify writing your own o/s, using unix, using something else, etc. -- bill davidsen (davidsen@crdos1.crd.GE.COM -or- uunet!crdgw1!crdos1!davidsen) GE Corp R&D Center, Information Systems Operation, tech support group Moderator comp.binaries.ibm.pc and 386-users digest. "I admit I am predjudiced, I can't stand bigots." -me
wailes@en.ecn.purdue.edu (Tom S Wailes) (06/26/91)
In article <3457@crdos1.crd.ge.COM>, davidsen@crdos1.crd.ge.COM (Wm E Davidsen Jr) writes: > In article <1991Jun23.012644.12449@en.ecn.purdue.edu> wailes@en.ecn.purdue.edu (Tom S Wailes) writes: > > | What word size would be the best for use > | and why? > > Since the answer all depends on the definition of "best" I guess we're > going to waste several weeks talking at cross purposes. Why not restate > the question giving some hint what you're trying to maximize. > > You can go for max MIPS, MIPS/$ (with or without software cost > included), lowest power drain, or best paint job. You can specify > writing your own o/s, using unix, using something else, etc. > -- > bill davidsen (davidsen@crdos1.crd.GE.COM -or- uunet!crdgw1!crdos1!davidsen) The best architecture would be one that maximized Specs, or some other measure of performance related to actual code. One problem I have is defining this. What code runs on a multiprocessor massively parallel machine at an actual company or research center? Do actual users run dusty decks of Fortran originally written for vector type machines, or do they experiment by developing entirely new approaches taking into account that the guts are hundreds of individual 32-bit micros. Do the present operating systems limit our experience such that present use would be much, much different than future use? The end goal of my research is the development of the hardware architecture of future "Killer Micros." I confess that this goal is elusive, because the micro hosts are constantly changing and improving , but I am trying to attack the shortest job first. How big is big enough. It seems 32-bits is big enough, but what do I know, I'm not a user. Tom wailes@ecn.purdue.edu P.S. On a related topic, what does one do with virtual memory on a Killer Micro? Do you distribute memory among the processors or do you create a large banked shared memory? Caching and virtual memory support for the uniprocessor environment will most likely be intrinsic on a 32-bit micro. How does one use this built in feature on a massively parallel microprocessor based architecture? A shared memory would offer better utilization in my opinion, but then it would not be local. Who has experimented with this?