home *** CD-ROM | disk | FTP | other *** search
- Path: sparky!uunet!decwrl!infopiz!lupine!motcsd!starnet!jpp
- Newsgroups: comp.arch
- Subject: Re: BUSES
- Message-ID: <1992Jul28.180537.14749@StarConn.com>
- From: jpp@StarConn.com (John Pettitt)
- Date: 28 Jul 92 18:05:37 GMT
- References: <1992Jul23.092211.18462@nuscc.nus.sg> <1992Jul23.191927.1181@pcnntp.apple.com> <1992Jul27.191347.4485@ksmith.uucp>
- Organization: Starnet-Public Access UNIX--Los Altos, CA 415-949-3133
- Lines: 51
-
- keith@ksmith.uucp (Keith Smith) writes:
- >Is it not also true however, that the current disk devices can't even
- >swamp an ISA bus? The average desktop doesn't have multiple striped
- >disks and whatnot.
-
- >Even the new crop cannot swamp an EISA bus. Video may soon go another
- >route, we will see. Memory is already localized, as part of the CPU on
- >most products, with a SIMM pinout becomming a local memory bus of sorts.
-
- Not true - on a 486 multi processor system running UNIX (an extreme
- example I will admit) bus bandwidth for disk, network and serial I/O is
- a major concern.
-
- On a more normal level an ISA bus will peak out at around 3.3 MB a
- second half duplex (600 ns per transaction 2 bytes per). In the real
- world where things don't run 0 wait and you have to do read modify write
- transactions on byte registers things slow way down and just over 1Mb /
- second is about all you get without using bus mastering.
-
- EISA is a better proposition but even their the 33 MB/sec theorectical
- max is not attainable on any system I know of. ~10Mb sec is more
- normal.
-
- In paractice the pure data rate is often not the issue - you don't want
- to more more than a 2 megs a scond on a PC that often. However as most
- PC's either don't have write buffers or have at most 4 levels of
- buffering the main x86 CPU is held in wait state while the bus
- transaction finishes. This significantly impacts performance. For
- example a 486/25 Micro Channel system (ALR in this case) running 128
- serial ports at 38.4K baud (~5 mbits /sec) takes a 60% of the CPU. If
- you look further into this of that 60% some 2 thirds (40% of real time)
- is spent waiting for the bus to finish a transaction. Whilst in the
- real world people don't run all terminals at 38.4 100% duty you can see
- that bus speed does make a difference. N.B. these numbers are from a
- real live set of tests done during the design of a fault tolerant I/O
- solution.
-
- So waht to do about it. Well fistly keep things in balance - EISA is
- fine for most PC applications. Beyond that more write buffering or a
- separate I/O controller with high speed access to memory and it's own
- EISA or ISA bus may be the way to go (similar to the channel controller
- concept in mainframe systems).
-
- Regards
-
- John Pettitt I/O system designer (retired :-)
-
- --
- John Pettitt jpp@starconn.com
- Archer N81034 apple!starnet!jpp
- Me, say that, never: It's a forged posting! Fax: +1 415 949 2037
-