Path: utzoo!utgpu!news-server.csri.toronto.edu!clyde.concordia.ca!uunet!samsung!think!leander.think.com!ephraim
From: ephraim@leander.think.com (Ephraim Vishniac)
Newsgroups: comp.sys.mac.hardware
Subject: Re: Hard disk benchmarks for Mac IIfx?
Keywords: disk, fx
Message-ID: <36089@think.Think.COM>
Date: 4 May 90 13:15:44 GMT
References: <1800@ndmath.UUCP> <36069@think.Think.COM>
Sender: news@Think.COM
Reply-To: ephraim@think.com (Ephraim Vishniac)
Organization: Thinking Machines Corporation, Cambridge MA, USA
Lines: 36

In article <36069@think.Think.COM> ephraim@think.com (Ephraim Vishniac) writes:
>I just ran a simple test using SCSI Evaluator 1.02 on a IIfx.  My
>internal disk is an Apple (Quantum) 80MB disk; the external disk is a
>CDC Wren III.  Both disks are formatted with 1:1 interleaving.  I ran
>SCSI Evaluator's default tests, for reads only.  For small transfers
>(10K, 15K), the Apple 80MB disk was literally off the chart.  I
>suspect this means that the controller caches more than 15K, but less
>than 20K.  For 20K transfers and larger, the Apple disk did about
>6000kbps. 

After I wrote this yesterday, my memory suddenly engaged.  The Quantum
disks use a multi-way read-ahead cache.  That is, if you read a group
of sectors ending with X, the disk controller uses its idle time to
read sectors starting at X+1 into the cache.  If that's what you read
next, you win.  If not, you haven't lost anything because the disk was
just sitting there anyway. 

The number and length of read-ahead groups is configurable.  If you
suspect, for example, that you'll typically switch back and forth
between reading three different areas of the disk, you can divide the
64K cache three ways.  From the SCSI Evaluator results, I guess Apple
has divided it four ways, with 16K of look-ahead for the four most
recently read sectors. 

There are actually two benefits to this kind of caching.  Obviously,
the transfer of cached data is very quick because there's no
rotational latency.  Also, the number of seeks is reduced.  This is
extremely important in reducing working time for disk-bound
operations. It's also very difficult to judge the importance of it
from simple benchmarks.  There's nothing like running your own
application to judge the disks. 

Ephraim Vishniac    ephraim@think.com   ThinkingCorp@applelink.apple.com
 Thinking Machines Corporation / 245 First Street / Cambridge, MA 02142
        One of the flaws in the anarchic bopper society was
        the ease with which such crazed rumors could spread.