Path: utzoo!utgpu!news-server.csri.toronto.edu!cs.utexas.edu!samsung!olivea!orc!inews!iwarp.intel.com!gargoyle!ddsw1!karl From: karl@ddsw1.MCS.COM (Karl Denninger) Newsgroups: comp.sys.amiga Subject: Re: Re: Re: A3000UX Seems Fated (Kill file alert!) Summary: More discussion.... Message-ID: <1990Dec25.234322.836@ddsw1.MCS.COM> Date: 25 Dec 90 23:43:22 GMT References: <39228@nigel.ee.udel.edu> <1990Dec17.052316.19609@NCoast.ORG> <1990Dec24.220257.12208@ddsw1.MCS.COM> <1990Dec25.045537.13517@NCoast.ORG> Reply-To: karl@ddsw1.MCS.COM (Karl Denninger) Organization: Macro Computer Solutions, Inc., Wheeling, IL Lines: 194 In article <1990Dec25.045537.13517@NCoast.ORG> davewt@NCoast.ORG (David Wright) writes: >In article <1990Dec24.220257.12208@ddsw1.MCS.COM> karl@ddsw1.MCS.COM (Karl Denninger) writes: >>"Seldom" does not equal "not available" on the RAM expansion. Check with >>the 486 owners -- most of them can do 64MB on the motherboard, using 4MB >>SIMMS. No problem; plug and play. > Sure, but most people consider the '386 similar in performance to an >'030 (though I am not one of them. In real-world use I have found the '030 to >be faster). If you look at systems designed around the '040 you might also find >that. I am also sure it has something to do with the price and market. No one >would buy a '486 system that handled that much memory just to run DOS >software, it would most certainly be for Unix. I am also sure that if you are >using 4mb SIMMs (which aint cheap by far), you will be quite limited in your >memory configuration, as you useually are when you use a system (like the >Amiga 3000) that can handle 256k and 1mb chips. You either have a little >RAM, or a lot of RAM, without small increments inbetween. For 4mb chips I >would assume this would be even worse. Actually, this isn't that bad at all. With the systems I've sold and worked with, you can choose 1, 2, 4, 8, 16, 32, 48 or 64MB of RAM. That ought to be enough choices for nearly anyone. There ARE uses for 8MB+ in the DOS world -- Windows 3.0 in enhanced mode for example. And oh, by the way, Windows 3.0 is true multitasking AND protection of system memory space (ie: task partitioning) as well. '386 systems CAN equal the performance of an '030, depending on the memory architecture and design of the system as a whole. The main determinant is the cache arrangement; how well it's implemented and what method is used. Those differences don't tend to show up on benchmarks like the dhrystone, but they sure do under multiuser load. Properly designed 486 systems blow the '030 out of the water. The '040 is a reasonable match there, again dependant on the intelligence of the system design. Neither chip has an inherent advantage, unless you want to run MSDOS :-) . >>This is a non-issue. The ISA PC bus can pass 8MB of data per second. The >>best SCSI fixed disks we have available today can do 2MB/second. MOST disks >>cannot sustain that kind of transfer rate for long, even in synchronous >>mode. > Hogwash. It may be able to do BURSTS of that high, but I would like to >see an 8 Mhz bus do 8mb a sec for any amount of time, even if the controller >was taking over the machine to do it. Remember, some ISA PC's (this one for example) can clock at either 8, 10, 12 or 16Mhz. My current SCSI controller (Aha1542) can run comfortably at 50% bus utilization at an 8MHZ rate; this will max 2 SCSI drive transfer rates without trouble. If I'm willing to do some tweaking I can get perhaps 10-20% more, but this is more than the drives themselves can deliver, so I'm quite happy. In fact, I have to back off the controller transfer rate so as not to starve the drive read-ahead cache (which causes SCSI disconnects and thus slows down the transfer rate) If you shop motherboards carefully, and mess with the Aha1542 parameter block, you can get a 10Mhz transfer rate at 60% utilization without problems. >>SCSI has a maximum transfer rate of 5MB/sec. That is STILL well under the >>ISA bus machine's bus bandwidth. > You mean SCSI-I here, not SCSI-II. Yes, I mean affordable disk drives. :-) >>Since there is little or no reason to put anything ELSE on that bus other >>than perhaps video, you have 4X the needed processor <> disk bandwidth. >>Anyone who tries to sell you more bandwidth than you need is a thief -- of >>your money, that is. > Oh? Like maybe 64 port I/O boards? Or maybe even 16 port boards? Just >how many users are you trying to stick in there without using some kind of >I/O board in the ISA backplane? And all these other boards you are adding >are further cutting down on the available bandwidth of the ISA bus. Oh? Let's see..... 64 X 3840 X 2 bytes per second (38.4kbaud) = ~500 KB/sec, or not enough to even bother. It's lost in the noise and overhead. That's with a good (ie: Equinox) design that knows how to take the load off the processor. I have 24 ports in this system; with all 24 up, full duplex, 38.4k, it requires something like 5% of the processor to service the card. And a vanishingly small portion of ISA bus bandwidth. Interrupt-driven I/O on ANY system CPU will murder it. That's why you use intelligent I/O boards for these kinds of things. >Next, >if you are trying to run X on a VGA board, that will suck up a lot of the >ISA bus bandwidth right there, not to mention the main CPU time. Anyone who does this deserves what they get. Get a TIGA board if you want to do X windows; 14" screens aren't big enough anyway -- REGARDLESS of processor or system type. 17" is a minimum for reasonable utilization under "X". Try running Framemaker on a 14" screen sometime -- it's unusable. VGA is a cheap hack; it has reasonable resolution, fabulous color, but it isn't fast. No one has claimed it is/was. Get real video capability if you want it; the price isn't that unreasonable. Certainly cheaper on the ISA machine than equivalent capability on the Ami! (TIGA and i860 boards are just coming online for the Amigas) And for Gods' sake, get the frame buffer memory OUT of the system memory map and on a private bus where it belongs! >And finally, >have you ever used a SCSI system with more than one drive? Don't forget that >ALL the drives on a single host adapter share the same data port, and therefore >help to eat up any remaining room that might have been left. If you have 2 or >three SCSI devices, (best example would be a hard disk and a tape drive, >with the tape drive backing up the hard disk), it is EASY to hit the limit >of the ISA bus, even if the drives only do 2mb/sec. And as many of the better >drives have on-board caches, which do read-ahead, you are likely to get big >gushes of data coming in, from each device on the SCSI bus. Oh? This system, right here, has THREE Maxtor SCSI drives on it PLUS a SCSI tape. Sure, I can't max all three drives -- with one adapter anyway. With two adapters I can get damn close; Chantel makes the drivers which will allow this to work. I also have 24 high-speed serial ports on that same bus, and ethernet. The bus is STILL idle, at maximum registered load, more than 20% of the time. And I'm running ISA on this system; I don't have EISA. In fact, if I crank up the Adaptec the drive caches starve and cause SCSI bus disconnects. That is, the faster transfer rates on the controller translate to SLOWER real-world throughput! The fact of the matter is that you can't get anywhere near 2MB/second out of today's SCSI drives. A little over 1MB/sec is the best SCSI transfer rate I've measured on ANY system, and that was on a MIPS 3260 server. Cheating by repeatedly reading the same sector or track doesn't count -- I'm talking about reading 30-40MB data areas sequentially, so as to take into account head movement and settling time. When you start doing REAL I/O, with head movement all over the disk, you're not going to achieve anywhere near those rates with ANY disk drive. In the >real world<, 1MB/sec of sustained transfer rate per spindle is FANTASTIC. If you want more, you have to go to IPI or IPI-2 technology, which can do roughly double that (2.25MB/sec in real transfer rate). But those are 8" drives.... and much more money! That's ok though -- I can get more throughput with the 1542B than my $35K Sun workstation at work does under equivalent load. I didn't spend anywhere near as much money as Sun did either :-). >>Note that "real" controllers for the ISA bus, those which do bus mastering >>(AHA1542B for example) are here and do work. They take nearly all the load >>off the processor, do intelligent seek ordering, overlapped command/data >>requests, and all kinds of other wonderful things. > Agreed. And using a caching controller can help a lot too. A caching controller is usually a lose on a Unix machine. Put the memory on the system board, and crank the cache buffers. MOST of the time that's a win. The sole exception today seems to be when you have a 16MB hard RAM limit, and can use it all for user processes - then the cache controller gives you additional "buffer pool". Of course, with either of these strategies, a UPS is a must -- otherwise one power hit and the drive's data turns into soggy bread (which will make you MOST unhappy!). >>Today EISA is a waste of money, as is MCA. ISA is more bus than you can use >>for disk I/O. Hell, I can do better with an ISA PC than most workstations >>in disk performance with a little judicious tuning of disk parameters! > Depends on who you are selling to. To Joe Blow off the street, that's >true. But this thread was about the performance of ISA systems as Unix >servers/workstations. Jow Blow doesn't do this kind of work, and certainly >wouldn't pay the >$6000 for an entry level 16 user system. The only way ISA is >"more than you can use" is if you have ONE device attached to it, and don't >expect to do lots of large-scale data transfer (like swaps). Nonsense. 486 ISA systems make damn fine Unix workstations. The problem with servers is in the Ethernet performance, and that problem lies in the driver and board maker's camp -- not the ISA bus. Ethernet can pass, at a maximum, 1MB (megabyte) per second. Not a bit more. For high-speed graphics, get a real graphics card, and you'll have what you need. Don't try to make the processor do work it wasn't designed for through an inferior interface (VGA). NEITHER the Amiga 3000 or the ISA machine will make a good server. For that you need a real system bus, such as VME. You need real ethernet cards which can sustain high throughput (none of the ISA or Ami ones can do this at the protocol level). You need real drivers, which can use the power you have in the cabinet, and enough CPU to drive the boards to their limits. You need an operating system which is tuned specifically for this purpose. There's more needed too, but you get the idea.... Sure, I can make a reasonable server out of a Sun Sparcstation, a Mips Magnum, an ISA machine, and probably even an Amiga 3000UX. The question is how much throughput do you really expect to have, and how well will that system coexist with the rest of your network. I know where the limits are on the ISA system, the MIPS series, and the Suns. I have no idea on the 3000UX -- but I'll bet ethernet performance will be one of the sore spots. Hell, ISA Ethernet is faster than Ethernet on MCA systems! -- Karl Denninger (karl@ddsw1.MCS.COM, !ddsw1!karl) Public Access Data Line: [+1 708 808-7300], Voice: [+1 708 808-7200] Macro Computer Solutions, Inc. "Quality Solutions at a Fair Price"