Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Path: utzoo!utgpu!water!watmath!clyde!rutgers!rochester!bbn!husc6!mit-eddie!ll-xn!ames!amdahl!amdcad!rpw3
From: rpw3@amdcad.UUCP
Newsgroups: comp.dcom.lans,comp.sys.ibm.pc
Subject: Re: PC LAN Comparison
Message-ID: <19194@amdcad.AMD.COM>
Date: Thu, 19-Nov-87 01:42:40 EST
Article-I.D.: amdcad.19194
Posted: Thu Nov 19 01:42:40 1987
Date-Received: Sat, 21-Nov-87 14:07:15 EST
References: <2070@killer.UUCP> <1020@kodak.UUCP> <155@tic.UUCP>
Reply-To: rpw3@amdcad.UUCP ()
Organization: [Consultant] San Mateo, CA
Lines: 41
Keywords: IBM PC, Novell, hardware
Xref: utgpu comp.dcom.lans:844 comp.sys.ibm.pc:8810
Summary: DMA isn't *always* "better"

In article <155@tic.UUCP> ruiu@tic.UUCP (Dragos Ruiu) writes:
+---------------
| I too agree that the Novell report was an unusually useful response...
| However, I had great difficulty believing their hardware section, which
| repeatedly says that DMA interface boards are slower than others. Unless
| there is something I don't know, DMA is much faster than interrupt driven
| hardware.
+---------------

Surprisingly, not always! I have seen many cases over the years where a
clean-and-simple interrupt-driven design beat out a DMA design, especially
when the DMA hardware was clumsy, hard to set up and get started, and hard
to re-start. The PC's DMA chip loses on all counts in my book, but you can't
do DMA without it, since the bus doesn't let peripherals do "true DMA". (This
is partially fixed in the -AT, and even more so in the PS/2.)

DMA can also be a loser is "low latency" applications, where getting to
the data quickly is more important than moving bulk along, especially
when the hardware must be touched in "realtime" to make some protocol
work (as is often the case in 3270 bisync, for example).

"It ain't that simple" is one of the variants of Murphy's Law. The design
dimensions of low latency, throughput, and CPU overhead are ALWAYS mutually
antagonistic, if not exclusive. Using DMA well requires a bus design that
supports multiple bus masters, with low overhead switching from one to another.
Few low-cost computers have such a bus.

I don't know whether Novell is simply touting THEIR particular design,
or whether they've done a careful analysis and decided that DMA just
isn't worth it for a typical LAN controller (provided you have sufficient
on-board buffer and a simple way to map the buffer pages into memory).
I do know they COULD be right, based on situations I've seen before...


Rob Warnock
Systems Architecture Consultant

UUCP:	  {amdcad,fortune,sun,attmail}!redwood!rpw3
ATTmail:  !rpw3
DDD:	  (415)572-2607
USPS:	  627 26th Ave, San Mateo, CA  94403