From: utzoo!henry
Newsgroups: net.micro,net.unix-wizards,net.works
Title: shared file-server == time-sharing again 
Article-I.D.: utzoo.2837
Posted: Mon Feb 28 15:18:29 1983
Received: Mon Feb 28 15:18:29 1983

People who are constructing workstation networks with diskless machines
and shared disk servers should carefully consider two facts:

	1. Unix shows a very strong tendency to run disk-bound.  It would
	seem reasonable to assume that any sophisticated system will share
	this characteristic, in the absence of convincing evidence to the
	contrary.

	2. A shared file server will show the same sort of degradation
	of response with increasing load as a time-sharing machine does:
	the same mechanisms (contention for disk heads, etc.) are at work.

Considering that accessing a file over a network is not as fast as getting
to it on a (well-designed) local disk, this does not augur all that well
for diskless workstations.  Nobody disputes the usefulness of shared file
servers for bulk storage, but the performance implications of doing ALL
your disk i/o that way clearly need careful consideration.

To give credit where it is due:  as far as I know, the first person to
realize the Unix performance implications of shared disk servers (as opposed
to the general usefulness of local disks for performance) was Tom Duff of
Lucasfilm's computer graphics lab.  His prediction has been confirmed
there, and they are now planning to put a local disk on each of their SUNs
for exactly this reason.  Tom recently lost the local disk on his SUN
temporarily, and says that the loss in performance is considerable.  His
estimate is that it takes three times as long to get to a file over a 10-Mb
Ethernet as it does on a local disk (with the same type of disk in both
places, I think).

					Henry Spencer
					U of Toronto