Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP Path: utzoo!utgpu!water!watmath!clyde!cbosgd!ihnp4!ptsfa!ames!husc6!uwvax!astroatc!johnw From: johnw@astroatc.UUCP Newsgroups: comp.arch Subject: Re: chewing up mips with graphics [parallel computing] Message-ID: <350@astroatc.UUCP> Date: Mon, 6-Jul-87 23:08:43 EDT Article-I.D.: astroatc.350 Posted: Mon Jul 6 23:08:43 1987 Date-Received: Thu, 9-Jul-87 04:39:32 EDT References: <8270@amdahl.amdahl.com> <359@rocky2.UUCP> <338@astroatc.UUCP> <1683@umn-cs.UUCP> Reply-To: johnw@astroatc.UUCP (John F. Wardale) Organization: Astronautics Technology Cntr, Madison, WI Lines: 65 Keywords: parrallel computing; performance; vector computing Summary: New algorithms good, but costly to port! ------- Comment on using up MIPS: a 100 mip or 1gip or 10gip ... workstation *IS* useful, even if its utilization is 1% .1% or .01% as long as it gives faster *RESPONCE* ... Any time you do something, the faster its done, the better! PS: I said useful, not necessarily cost-effective. =-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-= In article <1683@umn-cs.UUCP> herndon@umn-cs.UUCP (Robert Herndon) writes: >In article <338@astroatc.UUCP>, I wrote what prompted Robert. > Parallelize that, no. Parallelize the same problem, which presumably >is "tell me how projectile X under conditions Y will fall", is >entirely a different question, and may be parallelizable, >depending on the differential equation involved. Just as ...example of... [deleted for brevity]Your garden variety scanner is serial, but a fancy-new algorithm called "a Mealy Machine" is much better. > If you mean "ALL" problems, then I'd have to agree with you. >There certainly are problems that seem to require a serial >solution strategy. BUT: you must differentiate between >"a problem" and "code to solve a problem". They are NOT identical, >and an algorithm change makes it a whole new ball game. New methods are great for new codes, but given the amount of code in the world, and the number of new machines comming out, how much effort do you REALLY think goes into each conversion ??? I'd have a LOT more faith in your ideas if you had a parallel "guru" program that could identfy all your "poor" algorithms and fix them for you. (Maybe it could just say stuff like: xxx looks like a scanner. If you convert this, you could get 10:1 improvements, see file /usr/lib/guru/algorithm-examples/mealy for details.) (slightly :-) Programs run on Cray-cost machines can/will/are converted and/or fixed to run well, but most "users" are more likely to do simple "ports". Resatating this idea, If the machine is harder [arbitrarly comparison] to program, then it will not be accepted as well (it will have lower sales) ---------- I think you intentionally chose the "scanner" example so you could claim that its in the code distributed with the system, and will therefore be in all the vendor-supplied (system) programs, but scanning constitues a very *SMALL* fraction of the total CPU time, so these sort of algorithmic improvements must be applied in a wholesale manner to all the sub-parts of each program. That will (99.9% likely) be VERY expensive!! John W - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - Name: John F. Wardale UUCP: ... {seismo | harvard | ihnp4} !uwvax!astroatc!johnw arpa: astroatc!johnw@rsch.wisc.edu snail: 5800 Cottage Gr. Rd. ;;; Madison WI 53716 audio: 608-221-9001 eXt 110 To err is human, to really foul up world news requires the net!