Brian inquires regarding:
> The three main contenders are:
>
> Sparc 20/Solaris 2.5
>
> IBM C10/AIX 3.2.5
>
> IBM C20/AIX 4.1.4
>
> (ultrasparc)
The umich.edu cell mostly uses C10's running 3.2.5 as
fileservers at present. C10's are fairly decent machines
in a compact package; AIX 3.2.5 is a fairly mature OS
with few surprises. The most significant worry I
can see with this might be price.
C20's are definitely faster. I am not sure they're
enough faster to justify the price. AIX 4.1.4 is
newer technology. The major advances I know of are
(1) multi-processor support, and (2) more advanced
filesystem with block fragmentation. (1) will
not help straight fileserver performance at all.
It *may* help backups, vos moves, or if you use it
as a database server. (2) is a definite help
for disk usage, especially if you have many small
files in your setup (such as .o's.) This may
be reason enough to go with AIX 4.1 in a fileserver.
In a client workstation, AIX 4.1 may not be so
safe a bet; it took Transarc about a year to
produce a stable cache manager for solaris, and
AIX 4.1 could pose an equal challenge for them.
Sparc 20's are definitely pretty hot machines. They may
well be a cheaper bang for the buck as well.
On the other hand, solaris is more of a challenge
to program and administer. (My latest favorite:
gettimeofday(2)). For a fileserver, solaris
may well be a reasonable choice, although I haven't any
personal experience with them. It definitely took transarc
a while to get the AFS cache manager stable on solaris.
I'd expect to spend a bit of time collecting all the
right patches from Sun, & the right version of the
cache manager from transarc, to get a stable workstation.
The sparc20 seems to come with really weird internal
drives with funny connectors & plastic brackets.
I suppose it doesn't matter... On the other hand,
I'm not sure that the sparc20 is quite as expandable
as the C10 & C20.
The same comments regarding CPU on the C20
should apply to the ultrasparc. It's likely to make
the most difference for backups & the database
servers, & the least for a fileserver.
The most persuasive reason to get the C20 & ultrasparc
may be this: if you envision your cell growing, & adding
extra servers, you'll want to get lots of the same kind
of machine. Having lots all the same makes administration
easier now, & makes it easier to cannibalize parts in the
future. Therefore, get the thing with the best
future now, so you can get more later. If you do end
up with a heterogeneous environment; it's advantageous to make
the database servers as fast as possible, & it's advantageous to
make the database server with the lowest IP address fastest.
Besides CPU speed, I/O controllers, network latency,
& RAM size can also influence performance. The
network path, client workstation configuration,
administrative choices & user workload will also
all strongly influence what you will see. Don't
rely on industry benchmarks or conventional (even expert)
wisdom to predict fileserver performance. Try one
as a fileserver and measure it. On the plus side,
your users will never notice a difference of speed
that is less than a factor of 2, & of course
will never know what they haven't seen. Price,
simplicity, and reliability may be the most important
things to start with.
Off-hand, my feelings would be:
C10/3.2.5 if reliability > price, performance
if performance:
C20/4.1.4 esp. for database servers
if price:
sparc20/2.5 for fileservers
In all cases, 32M ram.
The sparc20 may be just as fast as the C20. Price
depends on what deal you can make with your vendor.
I don't have any really scientific facts to back
up my feelings, so count 'em for what they're worth.
-Marcus Watts
UM ITD PD&D Umich Systems Group
