Le Jeudi 4 Juillet 2002 19:41, C. Klein a écrit :
> Hi,
>
> take a look at this:
> .............
> ~/bin> uptime
> 7:32pm up 50 days, 4:10, 7 users, load average: 0.01, 0.09, 0.06
> .............
>
>
> That is the front end of our linux cluster, a bunch of athlons with a
> dual PIII front end machine. This all came for maybe 15000 USD and the
> performance is extremely good. The uptimes of the compute nodes are
> equally high. So do not worry about stability.
>
> The price/performance ratio is much much better for linux clusters than
> for SGI systems.
>
> I do not have a lot of time to write, but there is one very important
> issue: Once you get your cluster hardware, check it thoroughly. Check if
> all the hardware meets the specification (amounts of RAM, NIC types
> etc.). I had some surprises with that. And even more important: Do test
> runs of your software, if possible in parallel, for maybe a week. I had
> to send back 10 dual PIIIs because they would give unexplainable
> crashes!!!!! Do not pay the bill before everything is up and running...
> Take care not to buy cheap NICs.
>
> Best regards,
>
> Chris
I agree with Chris about stability and performance (you can even get athlons
instead of PII/PIV, you will get a much more efficiency on your calculation:
for example amber is 30% rapidlyer on an athlon 1.0Ghz than a 1.0Ghz pentium).
The only problem you will need to solve is the definition of your needs :
- first; will you use your cluster as a superparallel computer ?
in that case for example, you will not get a good speedup with explicit water
in sander, but a good one in GB. For PME on, you will better dedicate one
node per person.
- second: do your programs run under linux ? you talked about some virtual
screening, and if you think of Accelrys'product, they do not run under linux
for the moment.
Stef
Received on Fri Jul 05 2002 - 01:48:03 PDT
