Path: utzoo!utgpu!news-server.csri.toronto.edu!cs.utexas.edu!usc!samsung!sol.ctr.columbia.edu!ira.uka.de!smurf!flatlin!bad
From: bad@flatlin.ka.sub.org (Christoph Badura)
Newsgroups: comp.os.minix
Subject: Re: 386 dx vs. sx
Message-ID: <1705@flatlin.ka.sub.org>
Date: 10 Dec 90 21:57:19 GMT
References: <38578@nigel.ee.udel.edu>
Organization: Guru Systems, Karlsruhe, Germany
Lines: 57

In <38578@nigel.ee.udel.edu> KPURCELL@liverpool.ac.uk (Kevin Purcell) writes:

>So from this point of view, is the how does the difference in bus bandwidth
>of a dx compared to an sx affect the overall performance of a multi-user
>system.

Most probably not much. With normal MFM, RLL, and ESDI controllers the
cpu has to copy the data to and from the controllers on-board buffer.
With an AT-style backplane you get a 16-bit data path to the
controller in the BEST CASE. Because of some critical signal timing
constraints you may not get a 16-bit transfer to/from the controller,
*even* if the controller supports it. Furthermore you may suffer some
wait states when accessing the controller. In this case a faster cpu
won't give you more performance. It normally doesn't either since the
IO-bus speed is fixed at 8 or 10 MHz.

Installing a busmastering controller such as the Adaptec AHA154x or
Western Digital WD7000FASST SCSI controllers will gives you a *real*
performance boost and it will free a lot of cpu cycles for real work.
These controllers have a Transferrate between main memory an
controller of 5 MByte/sec and (depending on the motherboard) higher.
My disk throughput raised from somewhere around 90KB/sec with a 1:1
interleave MFM controller to about 300-400KB/sec with an AHA1542. This
is a 16MHz SX running Xenix.

>And a second question -- how does a cache on a 386dx affect the performance
>of Minix relative to a uncached 386dx. I have heard it said that (again
>with a real Unix system on a 386) a 20MHz cached 386dx will easily beat an
>uncached 25MHz 386dx.

That's right. In fact my zero wait-state SX compares favourable
against 20MHz DXen without cache (and normally some wait states) even
with the MFM controller.

>Finally, what affect is the hard disk system (disk + controller + driver)
>performance have on Minix. Are you better buying a better disk and controller
>and settling for a 386sx?

While faster disks give higher throughput, the real bottleneck is in
the file system code. The controllers have an upper limit on the
number of requests per second they can handle. Increasing the request
size has a dramatic effect on disk performancs. Typically, doubling the
request size doubles the throughput. So if someone has some patches to
FS, to read multiple continguos blocks in one chunk... hint, hint.

>All of these questions should be answered for the two cases that most people
>encounter -- the single-user case and the multi-user case.

Well, it depends. 1/2 :-)

I think currently FS is the main limiting factor regarding disk bandwith.


	chris
-- 
SVR4 is an adventure; if you win		 |	Christoph Badura 
 you find you're playing VMS. -- Richard O'Keefe |	bad@flatlin.ka.sub.org