Path: utzoo!utgpu!jarvis.csri.toronto.edu!mailrus!uflorida!haven!mimsy!fnord.umiacs.umd.edu!steve
From: steve@fnord.umiacs.umd.edu (Steven D. Miller)
Newsgroups: comp.unix.ultrix
Subject: Re: DECstation 3100
Message-ID: <17023@mimsy.UUCP>
Date: 20 Apr 89 16:18:13 GMT
References: <4706@decvax.dec.com> <88320@felix.UUCP> <1903@umbc3.UMBC.EDU>
Sender: nobody@mimsy.UUCP
Reply-To: steve@fnord.umiacs.umd.edu (Steven D. Miller)
Organization: UMIACS, Univ. of Maryland, College Park, MD 20742
Lines: 53

You're running a DS3100 diskless?  In most cases, that's like taking a good
runner and cutting his legs off, and then expecting him still to run fast...
(Caveat:  I don't even like running Sun-3/50s diskless, and let's face it, a
3/50 isn't in the same class as a DS3100.)

I don't think that DEC is wrong for selling diskless 8MB DS3100 systems.
There might actually be some people Out There who can use them -- I can
think of a secretary down the hall who would probably be thrilled to have
one.  The onus is on the purchaser to make sure that the system being
purchased will fulfill all the right requirements.  That's why I have mostly
12MB (or more) DS3100s, and that's why I have at least one RZ55 on each one.
(Sun will happily sell you a diskless Sun-4 of any variety, but would you
buy a diskless SPARCstation 390?)

I did notice that the performance difference between NFS mounting one's
binaries and running them off the local RZ55 is substantial enough to feel.
If you've got the older RZ55s without the readahead cache, you're not seeing
more than 400K/sec off the disk, and then you lose some of that through NFS;
it's no wonder.  I expect to see a big improvement in with the later RZ55s
(the ones that run more like 800K/sec), even across NFS.  (To check your rev
level on your RZ55s, do a 'scsi pb' at the monitor prompt, and if the rev
level comes back as 700 or less, join me in hoping for a FCO...)

I agree that the machine shouldn't crash under serious pounding, though.  It
should just slow down a lot.

	-Steve

P.S.:  I did a 'dd if=dxpsview of=/dev/null bs=8k', for the case where
the dxpsview binary (1368064 bytes) was on the local RZ55, and where it
was on a NFS-mounted filesystem.  The times I got were as follows:

	NFS:
	6 seconds (228K/sec)
	7 seconds (195K/sec)
	4 seconds (342K/sec, prolly out of buffer cache)
	average: 5.67 sec, 241K/sec

	RZ55:
	4 seconds (342K/sec) consistently for three tries

These numbers are totally off the cuff, but are, I think, somewhat
indicative; they indicate around a 40% speed loss for going to NFS.  So, if
we get 800K/sec out of a RZ55, we'd maybe see 480K/sec off the same disk via
NFS.  This ignores a lot of other factors, so maybe we won't see that,
either; at least it's a guess, that someone with a newer RZ55 could go
test out.  Mumble, mumble, mumble.

Spoken: Steve Miller    Domain: steve@mimsy.umd.edu    UUCP: uunet!mimsy!steve
Phone: +1-301-454-1808  USPS: UMIACS, Univ. of Maryland, College Park, MD 20742

Spoken: Steve Miller    Domain: steve@mimsy.umd.edu    UUCP: uunet!mimsy!steve
Phone: +1-301-454-1808  USPS: UMIACS, Univ. of Maryland, College Park, MD 20742