Return to BSD News archive
Xref: sserve comp.unix.pc-clone.32bit:4005 comp.unix.bsd:12371 comp.os.linux:52095 comp.unix.questions:37527 comp.os.mach:3148 comp.unix.solaris:5347 comp.windows.x:56902 Path: sserve!newshost.anu.edu.au!munnari.oz.au!constellation!osuunx.ucc.okstate.edu!moe.ksu.ksu.edu!vixen.cso.uiuc.edu!sdd.hp.com!spool.mu.edu!agate!tcsi.tcs.com!uunet!pipex!uknet!mcsun!Germany.EU.net!lemis!grog From: grog@lemis.uucp (Greg Lehey) Newsgroups: comp.unix.pc-clone.32bit,comp.unix.bsd,comp.os.linux,comp.unix.questions,comp.os.mach,comp.unix.solaris,comp.windows.x Subject: Re: Unix close for 486 - commens requested Message-ID: <3084@adagio.lemis.uucp> Date: 8 Aug 93 09:06:57 GMT References: <23r8kl$la4@pdq.coe.montana.edu> <CBAs9D.MH4@murdoch.acc.virginia.edu> <23rrf8$nrl@pdq.coe.montana.edu> Followup-To: comp.unix.pc-clone.32bit Distribution: inet Organization: LEMIS, 36325 Feldatal, Germany Lines: 105 In article <23rrf8$nrl@pdq.coe.montana.edu> osyjm@cs.montana.edu (Jaye Mathisen) writes: >In article <CBAs9D.MH4@murdoch.acc.virginia.edu>, >Jon Gefaell <jeg7e@livia.acs.Virginia.EDU> wrote: >>2.) You don't need accelerated video, 32M RAM, EISA, etc etc ad nauseum >> for what you've asked for. It's nice to have more and faster, but a >> lot less will produce very nicely for you. Especialy I have to wonder >> about the display recomendation given the stated application is for >> BATCH simulations. > >Well, there's a bit of a disagreement here. If the machines are only >batch, and there's no possiblity that they won't/can't be used for >X terminals or such, then you're right. But *if* you're going to buy >all those machines, and you have the opportunity to use them for computse >servers, and X, then accelerated video is the way to go. > >As to EISA vs ISA, it depends. If your simulation is so big that >you need more memory, then EISA is a better choice if you opt to run one >of the free unices, because of the 24bit addressing problem with the >Adaptec in an ISA box. The problem doesn't exist in the EISA box with >a EISA controller. EISA boxes aren't that much more than ISA now days >anyway. I don't recall offhand if BSDI implements bouncebuffers for the >154x series to support more than 16MB's of RAM on the ISA. The 16 MB limit has been discussed elsewhere - to the best of my knowledge, I'd just *love* to see some real numbers here. There's been a lot of hype about the performance improvements that EISA and Local Bus (particularly VESA) bring when compared to ISA, but I have never seen any numbers, and nobody has correlated these claims with the chip set in use. I have recently completed a reasonably comprehensive test of accelerated video boards under UNIX, and have found: 1. The price increment for a reasonably fast accelerated board (say the STB X-24, which runs an S3 801 and is about 15 times as fast as an ET4000-based board like the Diamond SpeedStar) is in the order of $100. Add the cost of a server (about $100 - $200 ) if your UNIX doesn't support accelerated boards (most System V don't, BSDI does). 2. With accelerated boards, the performance improvement through using EISA or Local Bus instead of ISA is hardly measurable. 3. The difference in motherboard chip set performance can more than offset the performance improvement of an EISA or Local Bus board. In my particular test, I compared S3 928 and CL5426 chipsets (like Elsa Winner 1000/#9 GXE and Genoa 8500 respectively) running under ISA, EISA and VESA local bus. The VL bus results were (slightly) *worse* than the ISA results. Running the test with the ISA board in the VL bus motherboard, I got results which were worse than in the vanilla ISA motherboard: obviously there is something wrong with the VL bus board. But nobody talks about relative motherboard performance, just these buzzwords EISA, ISA and VL Bus. I'll get round to more details later (maybe), but here are some orders of magnitude, measured on a 486DX/2-66 with 16 MB of memory and running SVR4.2: board bits/pixel line fill blt text arc cmplx xstones Elsa Winner 1000 (S3 928, 2 MB): EISA 4 312938 142525 115081 307656 2251175 195098 193895 ISA 4 309830 134609 115879 292875 1935545 146209 184360 VL Bus 4 311791 136694 122223 289437 1745757 120980 183283 ISA/VL board 4 295064 135495 120990 284281 1785385 106359 178251 STB X-24 (S3 801, 1 MB) ISA 4 195786 89979 78486 206937 1476028 123790 126570 ATI Ultra Pro (Mach 32, 2 MB) ISA 8 339402 58934 49199 183562 4786756 116078 100635 Genoa 8500 (CL5426, 1 MB) VL Bus 8 149458 28331 24077 177375 1983207 53398 53053 ISA 8 120519 28505 23929 207625 1801266 53856 52676 Diamond SpeedStar (ET4000, 1 MB) ISA 8 41113 5113 2663 68062 547235 5882 7823 In each case, I have chosen the pixel depth (4 or 8 bits/16 or 256 colours) which gave the best performance for the board). 4. Compared to motherboard performance, server performance is much more significant. There's been a reasonable amount of flaming recently about the relative performance of Metro Link and PPC. I haven't tested these servers yet, but I have tested the SGCS server. I didn't quite get their claimed performance (missed it by about 5% :-), but the results I did get were higher than Metro Link or PPC claim. >If the stuff is I/O intensive and reads and writes a lot of data, there's >no comparison between ISA and EISA, the EISA box blows it away. > >I've done some fairly extensive testing with the free unices (not Linux >however), and BSDI, and there is no comparison between IDE and good SCSI, >the SCSI blows it away. How about publishing your results? Please keep personal mail down to reasonable proportions. Greg -- Greg Lehey | Tel: +49-6637-1488 LEMIS | Fax: +49-6637-1489 Schellnhausen 2, 36325 Feldatal, Germany