jay@metran.UUCP (Jay Ts) (03/23/91)
(This post is really on two topics: the size and speed of the 7000/8000 Aviion systems. Size is discussed momentarily, then speed is discussed in great volume :-) :-) > In article <1991Mar14.183029.11714@neon.Stanford.EDU> torrie@cs.stanford.edu (Evan Torrie) writes: > > > >> Can anyone from Data General elaborate on the > >>architecture/capabilities of these machines? I looked in the paper > >>this morning, but there was no report, and only the picture of the > >>pizza box and what looked like a 4 processor 88K system. > > >The 7000/8000 Aviions are quad-processor servers. Four 25 MHz 88000's > >yield 117 MIPS of performance. [...] > > >I don't think they fit into a pizza box. > > The CPU board apparently does; at least, in the ads, they'd fit one into > a pizza box. I assume all four 88Ks are on the board in question.... OK, I finally saw the ad. If you read closely, it says: Data General's AViiON 7000 and 8000 systems have 117 MIPS of mainframe power that fits in a pizza box! That's right. The brains of these next-generation Open Systems occupy the same space as your basic large pie. The picture is of a quad-processor motherboard stuffed into a cardboard pizza box, on which is printed the Aviion/DG logo and ... 1-800-DATA-GEN (WE DELIVER) Apparently, this is to underscore the fact that the systems are shipping now, and (?) without delay. Maybe this is a poke at Sun's infamous order backlog. Jerry Callen writes: > Squeezing 117 MIPS out of (4 x 25MHz = 100 million machine clock ticks per > second) using a non-superscaler chip is an interesting accomplishment. Of > course, since 1 MIPS is a number that bears no relationship to anything > other than the willingness of marketing people to lie, I guess it's a fair > claim. > Robert Bedichek writes: > Just saying "everybody lies, so do we" doesn't cut it. The number must > come from somewhere. Even outrageous numbers have some basis. What is > the basis for 117 MIPS from the new DG machine? > > And of course, these DG figures are aggregate MIPS, not single CPU > MIPS, so it is meaningless to compare a 40 MIPS SUN to a ?? MIPS DG > machine. If the load is single thread, the fastest SUNs, MIPSco > machines, and the IBM RS/6000 will win. Until you add another process :-) Any good multiprocessor support will farm work out to the processors AT LEAST at the process level. If you are running a program under X Window, you have plenty more processes than processors. You end up with an overall performance increases that is surprisingly linear with respect to the number of processors. As far as the future is concerned, work is being done to support multiprocessing at the thread level (fine-grained multiprocessing). When a version of UNIX that supports this comes out, a multi-processor Aviion will be a very nice thing to have! > Someone from DG, please help us understand your company's claims. I have been trying to figure this stuff out too. With the help of an Aviion workstation I have on loan, I may be able to help shed some light on this matter. First, here are Data General's claims, from marketing info about a year old: model: 300 310 ------------------- # processors 1 1 clock speed (MHz) 16.7 20 dhrystone 1.1 37543 45167 MIPS 17 20 (Both the 300 and 310 are single-processor desktop workstations.) Now, about the MIPS rating. It's no good trying to compare 88000 MIPS directly to SPARC, mips or RS/6000 MIPS, because the architectures and instruction sets are different. If a SPARC, for example takes 2 instructions to do what the 88000 can do in one, a direct comparison will make the SPARC look twice as fast. I assume most of you know about this... Most manufacturers measure MIPS as VAX MIPS, where 1 VAX MIP is the speed of a VAX 11/780. But doing what? Well, it is common to compare based on the dhrystone 1.1 benchmark. This method, of course, is full of holes. Is the VAX running VMS or UNIX? If it is running UNIX, it will be significantly slower... Just for play, take a look at the following figures. hardware -> VAX 11/780 386 PC (64k cache) Aviion 310 operating system -> UNIX V.2 ISC 386/ix 2.0.2 DG/UX 4.3 -------------------------------------------------------------------------- clock speed (MHz) ? 20 20 dhrystone 1.1 1562 7364 40286 MIPS 1.0 4.7 25 Before I get into this further, a few notes: 1. The Aviion is a very old model, one of the first made. 2. The dhrystone figure for the Aviion is without X Window running; if X is running, it gets only 39117. 3. The dhrystone figure for the VAX is from the C source code for the dhrystone program. The numbers for the PC and the Aviion were obtained today by me, by compiling the C source and running it on the machine/os referenced above. I calculated MIPS based relative to the VAX dhrystone figure. What else do I have to go by? 4. These numbers are pure BS, as far as I'm concerned. They are here for conversational purposes only. End of disclaimer. Now for the fun. First, I'd like to point out that no matter what else, the Aviion, running at the SAME CLOCK SPEED as my 386 PC, and with LESS CACHE, ran the dhrystone benchmark almost 5.5 times as fast! (Some of this is due to better C compiler and kernel, but that's part of the product, you know.) Another reason I listed my PC is to bring up the point that Everex in their ads claims their 20-MHz/64K-cache PC (with just slightly higher dhrystone numbers than mine) is 4.9 MIPS, which is right in line with the 4.7 I calculate for mine. This leads into the subject of whether or not the figure for the VAX is any good. Well, the VAX is running UNIX, but so are the other machines. It doesn't make much sense to compare a more efficient operating system, does it? What if we run the dhrystone program under the Aviion's System Control Monitor, without UNIX getting in the way??? :-) Now, for a couple more notes: 1. I am trusting the dhrystone figure for the VAX without any knowledge of where it came from (maybe Everex, and probably a number of other companies did too...). If anyone has a more reliable figure, I welcome you to make me look incredibly stupid and post it. In fact, Please! I'd like better numbers for my own purposes, and if this is what it takes to get them, I find the humiliation acceptable. 2. My computation of MIPS figures for the 386 and Aviion is based on the VAX UNIX dhrystone. I got 25 MIPS for a 20 MHz Aviion! Now that, to me, looks like a lot. Data General was claiming only 20, which considering the above, may look a lot more reasonable to you now. 3. Those figures are for their old products. Somehow, between then and now, they have found a performance improvement of 17% across the board for their entire line. Whether that is due to better hardware, more efficient kernel and/or C compiler, or fantasies of the marketing staff is beyond me. Now, on to those new quad-processor servers. If you add 17% to their old MIPS rating for a 25 MHz machine (25 MIPS), you get 29.25 MIPS. (They claim 29 MIPS for a single-processor 25MHz machine.) Multiply that by 4 for a quad-processor machine and you get 117. Is this starting to make sense yet? You tell me! Well, if nothing else, this explains why most everyone is drifting away from quoting MIPS figures. What we really want is SPECmarks, right? I have some figures from Data General quoting SPECmarks for their Aviions. Unfortunately, they are apparently put relative to a VAX 11/780! I don't understand SPECmarks well enough to understand what is going on here, but anyway, here they are: machine: 300 310 SPECmark: 8.7 10.2 Now, if anyone knows "official" SPECmark numbers for a VAX 11/780, please post them! Jay Ts, Director Metran Technology uunet!pdn!tscs!metran!jay
lindsay@gandalf.cs.cmu.edu (Donald Lindsay) (03/24/91)
In article <384@metran.UUCP> jay@metran.UUCP (Jay Ts) writes: >Now, if anyone knows "official" SPECmark numbers for a VAX 11/780, please >post them! Officially, exactly one SPECmark. I don't have a SPECmark figure for recent DG products, but the Motorola 8864SP recently got 15.2 SPECmarks. This is for the same processor and cache, at the same clock rate (25MHz). Note that compilers differ, from vendor to vendor and release to release. -- Don D.C.Lindsay .. temporarily at Carnegie Mellon Robotics
mash@mips.com (John Mashey) (03/25/91)
In article <384@metran.UUCP> jay@metran.UUCP (Jay Ts) writes: ... >I have some figures from Data General quoting SPECmarks for their Aviions. >Unfortunately, they are apparently put relative to a VAX 11/780! I don't >understand SPECmarks well enough to understand what is going on here, but >anyway, here they are: > >machine: 300 310 >SPECmark: 8.7 10.2 > >Now, if anyone knows "official" SPECmark numbers for a VAX 11/780, please >post them! Yes. a VAX 11/780, with current compilers, is 1 SPECmark.... Note that SPECmark is a geometric mean of 10 benchmarks, 4 integer and 6 FP. If you believe mips are a measure of integer performance, then SPECint (the mean of the 4 integer benchmarks) is a better metric than Dhrystone... much better. Every computer architect knows that Dhryste overstates performance relae to the VAX more than real programs. In some sense, SPECint is about as good an approximation to vax-mips as you c get for lots of machines. The highest SPECint I've seen from DG (25MHz. 32KB cache, DG 6200) was 15.3. I saw some numbers from Motorola (25Mhz, 128KB, 8864SP) as high as 18.3. Tus, depending on compilers, memory system and such, one would expect thruput equivalant to 4X (SPECint for 1 CPU) - (adjustment for MP memory conflicts, which can range from a little to a lot). Bottom line: I conjecture that the 117-mips number is based on Dhrystone, even if it is labeled mips, rather than dhrystone-mips. I'd also conjecture that that number is 1.75-2X higher than a SPECint-based rating would give.... -- -john mashey DISCLAIMER: <generic disclaimer, I speak for me only, etc> UUCP: mash@mips.com OR {ames,decwrl,prls,pyramid}!mips!mash DDD: 408-524-7015, 524-8253 or (main number) 408-720-1700 USPS: MIPS Computer Systems MS 1/05, 930 E. Arques, Sunnyvale, CA 94086