🚀 go-pugleaf

john@iastate.edu (John Hascall) writes in response to a posting from
	joe@grc.genroco.com (Joe Nordman)

	[...A discussion about some assumptions in respect to the
paper about why SCSI is better than IPI for NFS servers...]

>   Not that I doubt any of what was said, but it only seems fair
>   that people know that GENROCO makes IPI controllers for DECs...

	None of that invalidates what Joe said, either. If I recall
the paper, it was based on the observation that high NFS traffic
tends to look fairly random, which defeats IPI's lookahead and
other strategies, therefore, since SCSI disks are cheaper, they
are better. In short: "NFS traffic is a pain no matter what, so
using cheaper disks is a good bet" which is somewhat true, but
doesn't take into account cases where the NFS traffic is bursty,
consisting of clients that do a whole series of sequential reads
or writes - in that case IPI is going to win pretty nicely.

	I'm not sure I'd recommend IPI as *the* way of making
my NFS server faster. In order I'd go with asynchronous NFS
and a UPS (UPS' on servers are a good idea anyhow) or an NVRAM
board if I had money to burn. IPI's going to be a lot better
for database or other applications like imaging than it will
be for NFS. Since the original question was *NOT* about NFS,
but was about how to make news faster, IPI is still a pretty
decent option, not only because of the faster disks, but
because they're nice and big, too(*). There are some amazingly
hot RAID controllers for IPI made by CDC, which can access
some ludicrous amount of storage in the 60Gb range. Enough
for most news feeds unless you carry the aquaria and MUD
groups.

mjr.

(*) You can also do fun still like hot-swap and dual port
IPI disks, which you can only do with SCSI if you like
having your bus hang, crash, and burn.
-- 
                                                 "guns don't die. people do."