Fast disk reads desired

[Rich Salz]

We will probably be starting a research project in the area of
free-text retrieval.  Looking around shows that our major bottleneck
is going to be how quickly we can read in large amounts of data
(eventually, in the Gigabyte range).

I'm asking the net for any hints and kinks people have used to increase
disk *read* throughput on Unix.  We're at an exploratory phase, so I'm
open to any and all suggestions, from "increase the block size to 16K"
to "buy a Convex." For example, does anyone have any feel for what kind
of performance I can get if I open a raw disk on 4.[23]BSD and do
vread(2)'s?  (I will probably not need the Unix filesystem structure.)

I know a great deal depends on the controller -- any word out there on
what's "best" and "fastest"?  At the start of the project, the group
will be sharing the machine with the rest of the company, although they
will probably have one or two controllers and disks of their own.  I
expect they'll need multiple controllers because of the need for
multiple simultaneous reads.  The initial development machine will
probably be something like a Pyramid 98x or a Sequent 21000.

This brings me to my second question.  The group may eventually move up
to a UTS-class machine in two years.  Does anyone have any numbers on
the overall disk throughput of those class of machines?  Or, what other
machines are good IO pumps?

Please reply to me, and I will summarize for the net.

Thanks,
	/r$
-- 

----
Rich $alz	{mit-eddie, ihnp4, wjh12, cca, cbosgd, seismo}!mirror!rs
Mirror Systems	2067 Massachusetts Avenue  Cambridge, MA  02140
Telephone:	617-661-0777					"Hi, mom!"