Re: sun sparc64 ultra 5 cluster (worth ?)
> I'm now running an Ultra 5 (275Mhz/ 64Mo/ 4.7Go HD) with a Debian
> Testing.
>
> Our University is getting rid of about 10 U-5s. Just the CU+mouse
> +keyboard (they keep the screen).
>
> We're looking for some ideas to use them for fun or for the use of
> someone.
One option might be to donate them to Free Software projects, either
maintaining them locally so developers could use them as development
machines (there seemed to be some confusion RE:The Vancouver plans over
whether Debian/SPARC had enough SPARC development machines) or to
directly offer them to an open source OS, perhaps to Debian or one of
the BSDs.
> We thought about making a cluster with them.
>
> We'll begin it home first.
>
> 10 U5s, that needs energy. So that we'll first have to pay for some
> extension of the power supply of our home.
Hmmm... don't know where in the world you are but in the UK (which I'm
told has a reasonably standard electricity system) you might not need
to. They are rated at 286 Watts, althought I suspect they actually draw
more like 100, if that. With a 13A ring main at 240V you should make
that. YMMV.
> That's expensive. My first
> question is to know if it's worth to clusterise 10 U5s under Debian.
Ummm... worth in what sense? What problems are you looking to solve?
How often do you plan to run this system and for how long? How much
money are you willing to invest? How much are you paying for
electricy? Will you need extra cooling?
If you want to build a cluster as an example of how to do it, to teach
yourself and others and gain an edge on a few problems then probably
yes. If you are looking for a good money/compute power trade off then
it depends on how much priority you put on initial cost vs. running
cost. If initial cost is the big issue then it might well be, if
running cost is more important than start up, I'd be tempted to suggest
a blade server of some sort or maybe wait until the dual core opertons
drop in price and buy a dual / quad motherboard.
Picking the best system for a particular task is not simple, you may
find that for a particular task the interconnects or the disks are more
important than the processors.
> It could be used to loop-compile/loop-build packages for sparc64 arch
> (as far as I remember, there is a xBSD doing that...).
distcc might be what you're looking for.
> I guess the cluster wont be 10 times faster (for compiling big
> packages), but would you know "à peu près [1]" the gain?
>From http://distcc.samba.org/
<quote>
distcc is nearly linearly scalable for small numbers of machines:
Building Linux 2.4.19 on a single 1700MHz Pentium IV machine with distcc
0.15 takes 6 minutes, 45 seconds. Using distcc across three such
machines on a 100Mbps switch takes only 2 minutes, 30 seconds: 2.6x
faster. The (unreachable) theoretical maximum speedup is 3.0x, so in
this case distcc scales with 89% efficiency.
You don't need a lot of machines to benefit: a laptop and a single
desktop is much faster than a laptop alone
</quote>
> If we do it, We'll perform some tests home and then once it is validated
> by our admin it could be installed into the university datacenter, wich
> has enough bandwidth....
If you were willing to have it as an official Debian buildd or
development machine then I'm sure there would be people who'd be most
interested.
HTH
Cheers,
- Martin
--
Martin
inkubus@interalpha.co.uk
"Seasons change, things come to pass"
Reply to: