[Beowulf] Keeping the Athlon MP cluster limping along

[David Mathog]

> 
> > The S2466N is an ATX form factor, each one has one Athlon MP
> > 2200+ and 1 Gb of 2100 DDR RAM, a 40G ATA disk, a floppy
> > and a little PCI graphics card in a 2U case.  If I could
> > find a nice mobo/CPU combo for, oh, <$200 that could
> > replace the S2466N and Athlon MP, and still do ECC, then I'd
> 
> Why exactly did you buy the more expensive Athlon MPs and dual
> motherboards, and then use only 1 cpu per motherboard?


1.  We were hoping the price of the MP chips would fall dramatically,
allowing us to fill out the servers.  What happened instead is that
it leveled out at around $150/chip and then they disappeared. (The
same trick had worked once before for us, with
some Intel Pentium II 400s and ASUS motherboards, the second CPUs
were dirt cheap when we bought them a few years after the initial
system purchase and they extended the life of those workstations.) 

2.  We needed ECC - these nodes run continuously.  ECC for
Athlons was very hard to find at the time we purchased
these machines.


> 
> I believe the socket 754 and 939 Athlon 64s do not support ECC, while
> the socket 940 Athlon 64 and Opteron do.

It's hard to tell what the 754 boards support since many say they
will _accept_ ECC memory but they don't say that they can actually
_use_ it.

> Prices per cpu seem to start
> around $125 for socket 754 or 939, $190 for socket 940, and go up from
> there, so you're going to have a very hard time squeaking in under
> $300 for the CPU + motherboard that you want, never mind $200.  Under
> $400, that you could do.

I was thinking of a 754 option primarily - 70 for motherboard + 130
for Athlon 64 2800 CPU (more or less.)

> 
> A cheaper option would be to keep the Athlon MP and simply replace the
> motherboard.  The Athlon MP will work just fine in any motherboard
> taking the Athlon XP, but finding a non-dual motherboard that supports
> ECC might be tricky. 

Tricky is right - name _one_ Athlon XP motherboard that supports ECC.

Besides, it may well be that the CPUs are bad and the
motherboards are all right.  Not having any spare known
good CPUs or known good motherboards there's no way to
play mix and match to figure out which component is the problem.
Well, not unless we sacrifice one of the working nodes, and
I'm really hesitant to do that in case it's one of those horrible
situations where component A breaks component B, which would
result in us having 3 flakey nodes instead of 2.  In any case, if
the CPUs are bad they'll be around $150 to replace from a vendor
and the motherboard about $190 (somewhat less on Ebay, but that's not
how I want to buy components.)

Regards,


David Mathog
mathog at caltech.edu
Manager, Sequence Analysis Facility, Biology Division, Caltech