Re: [ccp4bb]: CCP4 computing platform

[Kevin Cowtan <cowtan@ysbl.york.ac.uk>]

***  For details on how to be removed from this list visit the  ***
***          CCP4 home page http://www.ccp4.ac.uk         ***

On Friday 09 Aug 2002 11:49 am, Juergen Bosch wrote:
> I would go for an Xserve from Apple and in case your're not in favour of OS
> X.2,
.....
> and a few benchmarks:
>
> http://www.xinet.com/benchmarks/benchmarks.2002/

These benchmarks, while helpful for the selection of a file-serving machine, 
are not so useful for evaluating a compute server.

For number crunching, particularly single-precision crystallographic code, a 
dual athlon or dual pentium 4 will easily beat any comparatively priced 
machine from Apple or SGI, especially if you use a good compiler. For double 
precision calcs, the SGI or Apple fare a bit less badly.

If you buy it from a white box manufacturer, you can probably get at least 
twice the performance at half the cost. (We've been buying dual Athlon 1800s 
with software raid and 2Gb ECC ram for about $1600).

BUT: if you are using a white-box manfacturer, you'll need to be careful about 
cooling. An aluminum tower or a server case with at least 3 fans is a good 
idea, and the best CPU fans you can get (they'll be noisy - if its not going 
in a machine room, downgrade the CPUs and look for quiet fans). In either 
case, avoid the top CPU (or 2) in the range. Don't accept the first machine 
until you have run 2 concurrent cpuburn-s for 24 hours without a crash: see 
http://users.ev1.net/~redelm/

And get a warrenty.

If you want to build a cluster, Linux boxes are now cheaper than the high 
speed networking to link them together. However, if you've got someone who 
can write a routing table, then you can just stick a 4-port ethernet in each 
machine and use a combination of short hop/long hop wiring to do away with 
the networking hardware completely.

-- 
Department of Chemistry, University of York, Heslington, York YO10 5DD