Path: utzoo!utgpu!utstat!jarvis.csri.toronto.edu!mailrus!ames!killer!texbell!uhnix1!sugar!karl From: karl@sugar.uu.net (Karl Lehenbauer) Newsgroups: comp.os.misc Subject: When you have to go as fast as possible (was Re: Re^2: Unix bigotry) Summary: I recant somewhat, sometimes too long is too long. Message-ID: <3482@sugar.uu.net> Date: 22 Feb 89 04:42:18 GMT References: <4434@freja.diku.dk> <5900004@hpfcdc.HP.COM> <3472@sugar.uu.net> Organization: Sugar Land Unix - Houston, TX Lines: 49 In an earlier article, I presented an airplane analogy and said that, if you want to go as fast as possible, you have to give up amenities. I went on to say, essentially, run Unix on your supercomputers to save money on software development, and a bit more which I'll mention later. Brian Jay Gould (gould@pilot.njin.net) replied that he agreed to a point but that performance to the supercomputer user is more important than the cost of the programmer. "In engineering and scientific applications that require a supercomputer, the breakthroughs (or lack of) generated by computational science outweigh the cost of the supercomputer, software, and support staff by an order of magnitude." It sounds like they should be able to easily afford more CPU cycles then, but anyway, Uncle. Sometimes the ultimate speed is required, and one must "pay any price" to achieve it. In the airplane analogy, if the task is air combat, a DC-10 will not suffice, no matter how comfortable it is. In the computer world this means stuff like assembler, microprogramming the array processor, custom hardware, minimalist operating systems, etc. I contend, though, that too often people resort to all sorts of weird hacks when nothing, a better algorithm, or a more powerful machine would have sufficed, and that often the ultimate cost of the hack is much greater than anyone realizes, and if the costs were known, the decision to buy more computer power (or whatever) might seem rather attractive in comparison. Note for example the success of high-level languages, which are almost universally acknowledged as being less efficient in number of CPU cycles consumed for amount of computing done compared to expert assembly hacking, yet most production supercomputer programs are written in FORTRAN, for all well-known advantages thereby gained. I had kind of a dodge in the part of my posting comparing the performance of a RISC workstation to a supercomputer in that I said "scalar MIPS." Although I think some amazingly high performance vector units will be appearing for workstations, supercomputers will continue to have much higher I/O performance. I still think that if a site's supercomputing resources are being spread too thinly, that is, they are being shared by a lot of users running a lot of different programs rather than running a small number of long, difficult to partition (or nonpartionable) programs, workstations may be a win. On the other hand, although the newly emerging workstations of today (and tomorrow) are pushing into certain of the performance capabilities of the much more expensive supercomputers, the newly emerging supercomputers (such as the Cray-3) are themselves vastly more powerful, and will probably reassert a large performance differential versus workstations. -- -- uunet!sugar!karl | "Everyone has a purpose in life. Perhaps yours is -- karl@sugar.uu.net | watching television." -- David Letterman -- Usenet BBS (713) 438-5018