Return to BSD News archive
Path: sserve!newshost.anu.edu.au!harbinger.cc.monash.edu.au!bunyip.cc.uq.oz.au!munnari.oz.au!spool.mu.edu!howland.reston.ans.net!news.cac.psu.edu!news.pop.psu.edu!hudson.lm.com!godot.cc.duq.edu!ddsw1!not-for-mail
From: chilton@MCS.COM (Christopher Hilton)
Newsgroups: comp.os.386bsd.questions
Subject: Re: Need HW/Drive purchase advice
Date: 16 Mar 1995 09:16:12 -0600
Organization: /usr/lib/news/organi[sz]ation
Lines: 52
Message-ID: <3k9kns$ibd@Mercury.mcs.com>
References: <shamrock-1502952240270001@192.0.2.1> <3j3j5m$n70@helena.MT.net> <shamrock-0903951552350001@192.0.2.1> <3k167k$eb7@bonnie.tcd-dresden.de>
NNTP-Posting-Host: mercury.mcs.com
In article <3k167k$eb7@bonnie.tcd-dresden.de>,
J Wunsch <joerg_wunsch@uriah.heep.sax.de> wrote:
>Lucky Green <shamrock@netcom.com> wrote:
>
>[ISA bus-mastering DMA devices need bounce buffers with RAM > 16 MB]
>
>>I had just about decided on a Adapect 1542. But now I hear there might be
>>problems using it on ISA with >16 MB. Do you think I'd better get a 1522?
>
>No. They are PIO (programmed I/O), just like an IDE, and hence they
>will also hog the CPU as IDE drives do. No good decision for a multi-
>processing o/s. The penalty for the bounce buffers (if you happen to
>have > 16 MB) is much less.
>
>If you're buying a new machine and think about more than 16 MB RAM, do
>also think about PCI and a PCI SCSI controller (e.g. NCR ...810), they
>are rather cheap and do a good job -- and the bus doesn't suffer from
>this ancient address space limit.
>
I second this. Bounce buffers are a cheap price to pay for the
performance you get from bus mastering controllers.
Are these PCI controllers bus-mastering? I've been looking into
picking up an AVA-2822 as a stopgap until I can get a P90 or K5 with
PCI. This is an AIC-6360 on a VLB card. E.g. a VLB 1522. The reports
that I've seen say that it gets by the 3 MB/s limit of the ISA bus
quite nicely. Performance of a Quantum Lightning 540S went from 2.8
MB/s to 5.2 MB/s just by changing the card.
I understand intimately what you're saying about the AIC-6360 being a
CPU hog in PIO mode. Way back when I wrote the scsi driver for Mark
Williams Coherent. When I added code to chain multiple sequential
requests into one SCSI request to the disk driver I tripled the
throughput of the AHA-1542/Maxtor combo that I had. When I tried the
same trick on the Seagate ST01 controller I only added 20 percent. The
difference was that the Adaptec is "fire and forget". You put together
a SCSI request and then the controller handles the rest. The
Seagate however required hand-holding for each phase of the SCSI
protocol. This left no cpu to queue and chain the requests so 4 times
out of five the disk driver only saw one request at a time.
C.
--
Christopher Sean Hilton E-mail: chilton@mcs.com
----------------------------------------------------------------------
ICBM address: | "Thus it is said if you know them and know
42 07 39 N/87 49 44 W | yourself, your victory will not be imperiled.
For PGP key finger: | If you know Heaven and you know Earth, your
chilton@mcs.com | victory will be complete." - Sun Tsu
----------------------------------------------------------------------