Path: utzoo!utgpu!news-server.csri.toronto.edu!rpi!usc!snorkelwacker.mit.edu!paperboy!meissner From: meissner@osf.org (Michael Meissner) Newsgroups: comp.sys.m88k Subject: Re: Data General Aviion 7000, 8000 Message-ID: Date: 25 Mar 91 21:34:14 GMT References: <1991Mar14.183029.11714@neon.Stanford.EDU> <1991Mar16.151948.14745@dg-rtp.dg.com> <15515@june.cs.washington.edu> <1991Mar22.143552.16204@webo.dg.com> Sender: news@OSF.ORG Distribution: comp Organization: Open Software Foundation Lines: 63 In-reply-to: lewine@cheshirecat.webo.dg.com's message of 22 Mar 91 14:35:52 GMT In article <1991Mar22.143552.16204@webo.dg.com> lewine@cheshirecat.webo.dg.com (Donald Lewine) writes: | In article <15515@june.cs.washington.edu>, robertb@cs.washington.edu (Robert Bedichek) writes: | |> Would someone from Data General give a serious explanation for the | |> claimed 117 MIPS figure for the new workstations? | |> | OK. The MIPS quoted are Dhrystone MIPS. They mean that you can | get 117 more Dhrystones per second from an AViiON 7000 than from | a VAX-11/780. I don't like this method of MIPS rating, however, | it is the way the press likes to quote MIPS. It has nothing to | do with how many native instructions are executed per second. | [In fact, the 1.00 MIPS VAX-11/780 executes 500,000 VAX instructions | per second.] I tend to cringe when the marketing droids do this, particularly since I am one of the guilty parties. When I worked for Data General, I had originally worked on the MV C compiler, and then switched over to the GNU C compiler. At one point in the development of GNU C, we came to a crisis, where the C compiler generated something like 26,000 dhrystones (number suspect), and all I could get GNU C to do was about 17,000 dhrystones at the time. Eventually, we looked at the code generated, and discovered that the compiler turned: strcpy (buffer, "123456789012345678901234567890"); into: __word_copy (buffer, "123456789012345678901234567890", 31); Ah, competitive benchmarking.... The compiler also has a special option to turn off this optimization. I wonder why.... Anyway, I changed GNU C to make strcpy and memcpy builtin functions, and got about 27,000 dhrystones. In real live code, there aren't that many calls of strcpy with a string literal as the second argument, but it is in the middle of the dhrystone loop. Memcpy calls on the other hand, do occur quite frequently, so making them builtin is a reasonable optimization. | A more meaningful metric is the SPEC benchmark. The AViiON gets | an overall SPECthrughput of 54. It does even better if you leave | out the double precision nuclear physics test. I agree whole heartily that relying on dhrystone or whetstone for your performance measurement is kind of useless these days. The last time I looked at the ads, DG was at least forthright in saying they were reporting dhrystone performance figures. | All of these performance claims are only important relative to | price. A basic quad processor with 16Mb of memory, 660 Mb disk, | 500 Mb tape is under $100K. A price per SPEC that is very hard | to beat. Somebody else has mentioned that 16M of memory is rather skimpy. I certainly wouldn't want to run anybody's quad-RISC processor in 16M of physical memory (660M of disk is probably on the low side too). -- Michael Meissner email: meissner@osf.org phone: 617-621-8861 Open Software Foundation, 11 Cambridge Center, Cambridge, MA, 02142 Considering the flames and intolerance, shouldn't USENET be spelled ABUSENET?