Path: utzoo!utgpu!jarvis.csri.toronto.edu!mailrus!uwm.edu!lll-winken!brutus.cs.uiuc.edu!apple!oliveb!pyramid!decwrl!hplabs!hp-sdd!ucsdhub!sdcsvax!network.ucsd.edu!ucsd!tut.cis.ohio-state.edu!pacific.mps.ohio-state.edu!gem.mps.ohio-state.edu!rpi!batcomputer!kahn From: kahn@batcomputer.tn.cornell.edu (Shahin Kahn) Newsgroups: comp.arch Subject: Re: 1000000x1000000 Matrix (was: linpack) Message-ID: <91n:u@batcomputer.tn.cornell.edu> Date: 23 Oct 89 19:11:45 GMT References: <9089@batcomputer.tn.cornell.edu> <46500082@uxe.cso.uiuc.edu> <9118@batcomputer.tn.cornell.edu> <36553@lll-winken.LLNL.GOV> Organization: Theory Center, Cornell U., Ithaca NY Lines: 81 In article <36553@lll-winken.LLNL.GOV> brooks@maddog.llnl.gov (Eugene Brooks) writes: >In article <9118@batcomputer.tn.cornell.edu> kahn@batcomputer.tn.cornell.edu (Shahin Kahn) writes: >>If you have lots of scalar code (or code that doesnt run very fast >>on a super), and if it can't be rewritten, all you are doing >>is announcing *your* departure from supercomputers. >The series of supercomputers made by Cray was for some time the fastest you >could get for either scalar or vector coding. LLNL in fact has preferred them >for their superior scalar and short vector performance. (These days, of course Well, yes. these days you dont always buy a vector machine for its scalar speed. The problem is really because of lack of modularity in some vector machines, that you hhave to get the vector whether or not you like it. There is a definite trend towards more modularity, where the vector is an option. In fact, I see no reason why a single scalar cant support more than one vector! >we are not so pure in our decision making. We prefer Cray machines for their >software compatibility with the Cray-1. Although we are starting to work on >the problem, we have not fully embraced the notion of portable operating Like I said, its a good idea to REWRITE the code. The LLNL problem is not a problem. If you are on the cutting edge of things, sometimes you get cuts! >to keep track of some of our dusty decks.) The Cyber 205 could provide more >performance on long vectors, so I guess by your reasoning that you would call >it a supercomputer and accuse anyone buying a Cray machine of departing from >the real supercomputers. I guess that world has decided to not use enough I will give you the benefit ot the doubt and not even respond to this. Surely you don't really think that's what I meant. I only used the YMP because you had used the XMP as an example. >REAL SUPERCOMUTERS, because CDC could not sell enough of them to keep the 205 >and its children on the market. Japanese machines blow the doors off the Cray >machines these days (scalar or vector), but CRI is not worried about this. The ETA machine was/is a great machine and absoluttely first-rate technology. The problem of ETA was not the machine or the architecture. It boiled down to management even though they had a lot of good people at all levels. It was too little too late, and a problemm of not having deep-enough pockets and a lack of beleief, in the business community, in long-term.... > Their real nightmares have KILLER MICROS in them. YOUR nightmares, maybe. > >JUST BECAUSE ONE FRIGGING COMPUTER HAS A HIGHER UPPER BOUND ON THE FLOATING >POINT RATE THAN ANOTHER DOES NOT MEAN THAT IT IS A BETTER COMPUTER, OR THAT >IT IS THE "REAL SUPERCOMPUTER". THE BOTTOM LINE IS HOW FAST DOES A MACHINE All I asked was to request everyone to stop using the single processor XMP as the definition. Indeed, a supercomputer has many definitions and all of them are relative to some base characteristic. As you well know. >PROCESSORS OPTIMIZED FOR SCALAR PERFORMANCE. BECAUSE OF THIS I AM ACCUSED >OF DEPARTING FROM SUPERCOMPUTING. THIS IS GIBBERISH. IF I FIND A MIMD MACHINE >WITH 100 "KILLER MICROS" COSTING THE SAME AS A YMP AND WHICH RUNS MY APPLICATION >100 TIMES FASTER, I HAVE SIMPLY REDEFINED THE NOTION OF SUPERCOMPUTER. > FOR MY APPLICATION BINGO! For Y O U R application! That's the whole point. Funny we should start from a general red-alert about the attack of killerr micros and end up where we all knew about: "It depends on the application". Yes, if your application is "pagemaker" you redefine supercomputing by going to a mac-ii-ci !! Now in this case, there is something to be said about highly parallel machines. Indeed, there is a lot of that happenning right here. My own experience in supercomputing started with loosely coupled systems and scalar-lloking codes. But it's too early for the supers to pack up and go home. WAY too early. Universities are usually several years ahead of the market and they haven't got it quite right yet. By the time micros become killers, they wont be micros anymore