cy@dbase.UUCP (Cy Shuster) (12/08/88)
I just received two Apple Mac II EtherTalk cards (part # M0405), and connected two machines just to quick test the speed improvement. Surprisingly, doing a finder copy of a 1.2 meg file via TOPS, we get an overall data transfer rate of only 23K bytes per second! Times 8, that's only 184 kilobaud! Are these boards running at 1 megabit, or ten? And even so, that's a tremen- dous bandwith loss for protocols! (I'm used to a 50% loss for, say, xmodem over async). Finder 6.1a2, TOPS 2.0, EtherTalk v1.1, thin cable, straight line between the machines, with terminators. --Cy--
amanda@lts.UUCP (Amanda Walker) (12/11/88)
Well, TOPS is not the world's greatest speed test for networks. Over an Ethernet, TOPS is a little faster than over LocalTalk, but that's about it. My guess is that since TOPS runs "in the background," it only looks at the network every so often. AppleShare, on the other hand, is about as fast over EtherTalk as a local SCSI hard disk. The conclusion: the speed you get depends as much or more on the software as it does on the hardware. -- Amanda Walker ...!uunet!lts!amanda / lts!amanda@uunet.uu.net InterCon, 11732 Bowman Green Drive, Reston, VA 22090 -- "The best way to predict the future is to invent it." -- N. Negroponte
tim@hoptoad.uucp (Tim Maroney) (12/13/88)
A non-dedicated 8MHz 68000 cannot keep up with an Ethernet, period. The Mac Ethernet cards use heavy buffering on board. Aside from this problem, every network protocol has a certain scheduling latency. With Mac TOPS, which uses a proprietary synchronous scheduling mechanism, this latency is acceptable for LocalTalk or whatever they're calling AppleBus these days, but becomes rather noticeable on a higher-speed LAN. It would be nice if this could be reduced, but it would require a major rewrite of Mac TOPS. Even then, my guesstimate is that the maximum factor of improvement that the Mac could support would be about six, as opposed to the three times improvement you get now on Ethernet. Things might be better on the SEx (the 68030 SE scheduled for the first half of 1989). And let's not even talk about FlashTalk; benchmarks done at TOPS showed that the speedup was far less than three times, and under some conditions things actually ran slower. Don't believe everything you read in ads and in the lapdog trade press. -- Tim Maroney, Consultant, Eclectic Software, sun!hoptoad!tim "Gangsters would kidnap my math teacher, Miss Albertine, and I'd track them down and kill them one by one until she was free, and then she'd break off her engagement with my sarcastic English teacher, Mr. Richardson, because she'd fallen hopelessly in love with her grim-faced and silent fourteen-year-old savior." -- Nite Owl, in WATCHMEN by Alan Moore
lwe@suntops.UUCP (Leonard W. Edmondson @ Sun Micro, TOPS, Alameda, CA) (12/14/88)
Doing a 1.2 meg file finder copy via TOPS from a Mac II to a Sun 3/280 we observed a speed of 70K bytes per second and 60K bytes per second in the other direction. Len Edmondson
brad@cayman.COM (Brad Parker) (12/14/88)
From article <6047@hoptoad.uucp>, by tim@hoptoad.uucp (Tim Maroney):
> A non-dedicated 8MHz 68000 cannot keep up with an Ethernet, period.
10 megabits per seconds / 8 = 1.2 megabytes per second.
Reading words gives 600,000 word reads per second. (assume we have hardware
which will read 16 bits into a register from the wire)
8Mhz cpu give 125ns per cycle;
move.w An@,An@+ is 12 cycles (2 reads, 1 write), 12 x 125ns = 1.5 usecs/word
Which means I can do 666,667 word copies per second.
Doesn't this mean I could hand copy the bytes off the wire and still keep
up? (assume the loop is unwound)
Or did I miss something? I often do my math wrong, but
(assume that the sender will send small bursts
of "packet trains" which will give me some time to process the bytes
in between "trains".)
-brad
--
Brad Parker
Cayman Systems, Inc. Cambridge, Ma. brad@cayman.comA.ALDERSON@MACBETH.STANFORD.EDU (Rich Alderson) (12/16/88)
You've missed something. Most ethernets aren't populated with just two hosts. The wire is going to be busy most of the time, so you have to read everything on the wire to see if you are missing anything. That sounds like a dedicated processor (of any type) to me. A dedicated 8MHz 68000 could JUST keep up. It isn't going to have time to do much in the way of processing. Rich Alderson Stanford University -------
ragge@nada.kth.se (Ragnar Sundblad) (12/18/88)
In article <12454674092.150.A.ALDERSON@MACBETH.STANFORD.EDU> A.ALDERSON@MACBETH.STANFORD.EDU (Rich Alderson) writes: >You've missed something. Most ethernets aren't populated with just two hosts. >The wire is going to be busy most of the time, so you have to read everything >on the wire to see if you are missing anything. That sounds like a dedicated An ethernet busy most of the time is pretty unusable. Actually, a load higher than ~35-40 percent is absolutely maximum for getting things through. But, of course, you can consider that as a busy ethernet. >processor (of any type) to me. >A dedicated 8MHz 68000 could JUST keep up. It isn't going to have time to do >much in the way of processing. Probably, but this task is mostly done by a special designed coprocessor, often taking care of both address filtering and moving packets between the network and the host computer's memory, such as the NS8390 (as on Apples/3Coms EtherTalk card), or the Intel 82586. >Rich Alderson >Stanford University Ragnad Sundblad.
ts@cup.portal.com (Tim W Smith) (12/18/88)
Good heavens, don't we at least get to assume some sort of Ethernet chip in these calculations? These should "watch the wire" for us and only cause an interrupt when a packet for us arrives. Anyway, existing EtherTalk cards use on board buffering. Apple's EtherTalk card has either 32k or 64k ( I forget which ). Dove's FastNet III has 64k. FastNet II has 256k of RAM on board. FastNet SCSI contains 512K. FastNet SE has 64k. I don't know what Kinetics uses ( I don't have access to any of their products, but I have EtherTalk and all the FastNet stuff ). Some of the Ethernet chips are pretty nice about buffering. For example, the AMD LANCE, which Dove uses, lets you give it a ring of pointers to buffers. The chip fills the buffers and then sets a bit that says that the CPU owns the buffer. When the CPU is done with the buffer, it resets the bit so the chip knows that it can fill the buffer again ( actually, it is the other way around: a set bit means that the chip has the buffer and a clear bit means that the CPU owns the buffer ). The same thing is done for transmitting. FastNet III, for instance, uses 32 1.5k buffers, so a burst of up to 32 packets can be received before the CPU has to do something about them. Tim Smith