Date: Wed, 18 Mar 1998 08:58:06 -0800 (PST) From: Hugh LaMaster <lamaster@george.arc.nasa.gov> To: freebsd-current@FreeBSD.ORG Cc: lamaster@george.arc.nasa.gov Subject: Re: Stream_d benchmark... Wow, there really are differences in hardware Message-ID: <199803181658.IAA02816@george.arc.nasa.gov> In-Reply-To: <199803181057.LAA01642@sos.freebsd.dk> from "Søren Schmidt" at Mar 18, 98 11:57:28 am
next in thread | previous in thread | raw e-mail | index | archive | help
Soeren Schmidt (sos@FreeBSD.org) wrote: > In reply to Jaye Mathisen who wrote: > > Hmm, Then I should be proud of my noname system (p6/200/128MB 72pEDO): > > Function Rate (MB/s) RMS time Min time Max time > Copy: 117.0286 0.2758 0.2734 0.2812 : > Triad: 125.3878 0.3917 0.3828 0.4219 > > So what ?? Like all benchmarks, there is the question of "pride". However, computer performance is also a major problem/interest, and, stream is particularly informative for a "toy benchmark". Your numbers seem very high for a Natoma board - what type of EDO are you using and what are your BIOS settings? Exactly what kind of "noname" board is it? If these numbers are correct, I want one. Have you run the c't ctcm benchmark on it also? That gives a nice profile of the different memory bandwidth numbers from L1 cache, L2 cache, and main memory. [Note also, stream sometimes doesn't get the clock HZ set properly. Are you sure it is correct with these numbers?] [Maybe you have the 45ns EDO? What are your leadoff timings?] I'm impressed, anyway. I want to know more. In particular, I might be better off with one of your "noname" configurations than a 333 MHz P - II for driving some fast network interfaces. Your numbers look more like one of the specially-built and tweaked expensive "server" Orion boards. > > I figured that within reason, most mb's would have similar performance, > > but I was wrong. First of all, it depends on the chipset. Does the Digital Prioris ZX6000 system use the Orion chipset? It appears so. Note also whether using FP DRAM or EDO, and, how fast the memory is (the BIOS settings for lead-off timings may be slightly more aggressive for slightly faster memory. > > All boxes are P6-200's, 256MB RAM (all RAM is 60ns FP as far as I know). > > > > Box 1 is a SuperMicro P6DNE: > > Function Rate (MB/s) RMS time Min time Max time > > Copy: 60.7395 0.2704 0.2634 0.2832 > > Triad: 71.1647 0.3494 0.3372 0.3565 Typical for Natoma with FP DRAM I would guess. > > > Box 2 is a Digital Prioris HX6000 > > Copy: 73.3551 0.2197 0.2181 0.2249 > > Triad: 77.4268 0.3108 0.3100 0.3122 Is this with EDO? > > Box 3 is a Digital Prioris ZX6000 > > Function Rate (MB/s) RMS time Min time Max time > > Copy: 84.8807 0.2018 0.1885 0.2834 > > Scale: 97.5461 0.1661 0.1640 0.1720 > > Add: 111.6549 0.2179 0.2149 0.2247 > > Triad: 100.9468 0.2659 0.2377 0.4237 > > > > > > Box 3 uses 256bit interleaved memory, rather than whatever the > > "standard" is. I assume that this is an Orion GX chipset board. Yes, it does have higher memory bandwidth than Natoma if properly configured. Yes, IMHO, that does make a significant difference in the real world. > > I thought it was just a marketing gimmick, but it seems to really > > make a difference. Orion chipset boards had kind of a negative press at first due to problems with PCI bandwidth, but, I assume (?) those were worked out long ago. In any case, the Orion chipset never achieved the mass-market commodity status of Natoma chipset. As a consequence, boards remained, and remain expensive, and now, the PPro200 is being phased out. I wouldn't mind having one, but, there are only a few, expensive boards out there, such as the American Megatrends Goliath, and PPro200 prices never dropped that much, I guess because Intel wanted to phase it out in favor of the P-II. > > Have to see if it helps on some worldstone's. It should help a little bit, although, the real benefit tends to be on user apps which stream through a lot of memory. The classical FP applications, of course, but also, things like digital video, image processing, graphics, etc. - anything which streams through 1-4 MB of memory repeatedly. And, network performance to/from userland, which usually requires at least one data copy. To Unsubscribe: send mail to majordomo@FreeBSD.org with "unsubscribe freebsd-current" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?199803181658.IAA02816>