Re: [ccp4bb]: CCP4 computing platform

[Anastassis Perrakis <perrakis@nki.nl>]

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

> For CCP4 computing and other crystallographic processing.
> Would you prefer a big number-crunching server (SGI origin300 type) or a
> collection of high-end PC's running linux (separate or as a beowulf-type
> machine), with perhaps some money left for a serious party.

I thought the following might be of general interest for a slow month like
August...
So, sorry for posting to everybody this answer.

Well, If I was getting a *big* machine I would not get an SGI *for
crystallography*.
Even the highest end processor is at least 2 times slower than a good
pentium 4.
Your choice for big machine must be an Alpha ES45 or so ...
Disclaimer: I am not working for Compaq but I am biased since i have an ES40

after 10 yrs of SGI's and its great.

The exact choice depends to the 'architecture' of your lab.
In our lab, we are two groups with a total of 20 people. Most of these
people
spend their time doing biochemistry and want to login to a computer and have

all software running and ready to solve a structure. Also some people can be
very
busy at times with refinement and need lots of CPU while others don't need
computing at all. So, for me, it was firslty very clear I need a
'centralised' system
with lots of CPU in one box (physically or virtually). Then our CPUs are in
one
box, share the same memory and most important they have our nice RAID disks
locally !!! The drawback of *any* system with CPU's in different boxes, is
that
you need memory in each box (but thats cheap) and all the disks are mounted
through NFS ( I presume you will use at least one machine as a disk server).

NFS is notorius of slowing down I/O intensive processes, i.e. data
processing.

To counteract all these arguments, any AMD box is as fast as an Alpha EV866
these days or faster, and guess what, it costs 1/5 of the price or so. So if
you can either
tolerate the slow NFS *OR* educate people to put data on local disks
and take care of their backups *AND* either you don't care about people
having transparent acces to all CPU's *OR* you think people can use
efficiently *ALL* boxes, *AND* management overhead is not an issue
(upgrade systems in all machines instead of one) go PC's (or actually go
for OSX Macs where you get your Unix and 'Microsoft' environments
in one box without rebootings etc etc).

My personal opinion in brief is, that in a lab of expert crystallographers
which
spend most time in computers and not the wet-lab and for groups less than
5-7 people
I would get PC's (OSX Macs since I am in love with my brand new i-Lamp at
home).
For any group bigger than that I would favour centralisation.

    Tassos

PS Note that I did not examine the 'Beowulf' solution or 'Queueing system'
idea.
I wonder if people use them and can comment. If you have the time to manage
them or you
have a good sysmanager they would be attractive. I have never had the time
to
look in to that and learn how to do it properly, so I never gave it a
chance.
I think they have a very serious management overhead - but thats a naive
view.
I seem to recall that a few people built Alpha DS20 racks and a queueing
system and
are very happy.