jxh@decwrl.dec.com (Jim Hickstein) (10/08/90)
I am planning my upgrade to SunOS 4.1 and I've got myself into oscillation
on a question of partitioning my disks on my poor old 3/180S. First, some
background:
3/180S (2) xy451 with (2) Fujitsu Eagles M2351A,
Sun/2 SCSI adapter (eek!*) with CDC Wren-VI.
This machine serves five diskless clients, but these have heavy loads on
them during the day, and /export/swap seems to be a major bottleneck. At
night, I run some huge compilations of hundreds of files (they take about
8 hours a night).
First, I have decided to put back the second XY451, so each Eagle will be
on its own controller, and to split swap and /export/swap across these two
spindles. I have also decided to do things like keeping the binaries of
my product and /export/swap near each other on each disk; putting the
working directories for the nighttime compilations out on the SCSI disk in
their own partitions to minimize excursions across that disk during the
day, and during the compilations themselves; sizing /export/swap more
tightly to avoid seeking across lots of space that's never used; stuff
like that.
The question arises, once I have tiled up the disks for all these things
that need to be kept together, whether I should put people (~user) on the
Eagles (which I imagine to be faster than the SCSI disk), or to put them
out on sd0 to get a typical user job spread across three spindles. So:
what is the actual throughput in a typical job mix for an Eagle on an
XY451 compared to a Wren-VI on this old SCSI adapter? I see that the
Eagle's average access time is 18ms, and I imagine the Wren's isn't far
away, but does that old SCSI adapter, or the CPU (a mere 2 MIPS) get in
the way? What is the most important factor? Are the 4-year-old Eagles,
in fact, faster than the brand-new Wren-VI? What about raw transfer rate?
How is the SCSI on that score, really? How should I configure my new 4.1
kernel to use the server's 12MB to the best advantage?
* - I suspect the Sun/2 SCSI adapter of all kinds of bad behavior on the
basis of my experience that one cannot rewind SCSI tapes on it without
blocking the entire machine in the kernel somewhere. It's kinda old, so I
suspect that Sun can't fix this in software. My SCSI tapes are now on a
3/60, where they do not exhibit this distressing behavior. And a new
adapter is $3K: forget it. Given that it is this old and crufty in this
manifestation, shouldn't I suspect it of being horribly slow on disk
operations, too? How much faster would it go if I popped for the extra
$3K? Anybody measured this transition?
I would really appreciate hearing from some of you (I know you're out
there) who are nursing these old servers along, trying to squeeze the last
drops of blood from the old turnips (while, naturally, trying to get
funding for that 4/490 and truckload of SPARCstations :-). I'm new at
this (really, I am), and I need all your good advice. Thanks very much.
Jim Hickstein, Teradyne/Attain, San Jose CA, (408) 434-0822 FAX -0252
jxh%teradyne.com@apple.com ...!{amdcad!teda,sun!teda,apple}!attain!jxh