Re: Workstation dedicated for FLUKA

From: Colin Paul Gloster <Colin_Paul_Gloster_at_ACM.org>
Date: Thu, 4 Aug 2011 18:46:19 +0000 (UTC)

Dear Dr. Cybulski,

There was a similar discussion on MCNP-forum in August 2010. At the
end of this email there is a response specific to the email from you
on the FLUKA list. I mentioned in the MCNP discussion:
|----------------------------------------------------------------------------------|
|"[..] |
| |
|In general, there are no generalities. Everything is different. |
| |
|The most effective of the ways not involving spending money to judge |
|the prospective improvements by replacing hardware, would be to try |
|MCNP on hardware such as what you are considering buying. Does a |
|friend or the workplace have such hardware which you could use for a |
|while? (Making sure that you delete MCNP from that hardware after |
|trying.) Make sure that you try a range of representative simulations: |
|slow bottlenecks can be different for different kinds of simulations |
|and though some may be improved by one type of hardware, the same type |
|of hardware could aggravate a bottleneck in another kind of |
|simulation. |
| |
|[..] |
| |
|An AMD/Intel machine with more than two cores would have more Level 2 |
|caches, so might be faster than a dual core machine. Having more cores |
|free unused by Windows may make them available to MCNP, giving you an |
|opportunity to have faster simulations. Then again, trying to divide |
|the work amongst different cores may slow it down to be worse than |
|using a single core. There are no generalities. |
| |
|Some very expensive machines (such as in the Intel Xeon line) even |
|have Level 3 caches. If your simulations are being slowed down by |
|cache misses, then a Level 3 cache would help. (Actually, a bigger |
|Level 1 cache would help more, but I am not aware of different sizes |
|of Level 1 caches currently on the market: but I have not checked.) If |
|you are not being slowed down by cache misses, then opting for a |
|larger cache would be a waste of money. [You may find it easier to |
|get L3 caches in 2011.] |
| |
|Similarly, if there is a lot of parallelism in your simulations, then |
|additional cores might be worthwhile (and that is MIGHT), but would |
|definitely be a waste of money if your simulations have no |
|parallelism. |
| |
|Changing from a 32-bit operating system to a 64-bit operating system |
|might actually slow things down (again: MIGHT), but would definitely |
|allow you to install more memory chips for bigger |
|simulations. However, you did not mention that simulations are |
|currently being restricted by being too small. [..] |
|[..] |
| |
|[..] |
| |
|Well, if your MCNP simulations are forcing [..the operating system] to use virtual|
|memory, then an increase in memory would definitely speed things |
|up. Otherwise, extra memory would definitely make no difference to the |
|speed. (Also note that you can buy different speeds of memory chips.) |
|However, in either case, extra memory would allow you to run bigger |
|simulations. |
| |
|There are so many factors to consider, that empirically trying MCNP |
|out on different machines would be useful. However, you can also check |
|some things on the machine(s) which you already have... |
|To determine whether buying more memory (of the same type of memory |
|which you are currently using) would speed up MCNP: [..try commands |
|such as |
| top |
|or |
| ps -p IDENTIFIER --format pid,%cpu,rss,vsize,bsdtime,args |
|]. |
|If close to 100% of the memory is being used, then more memory would |
|speed up everything. |
| |
|To determine whether you are suffering from cache misses, use a |
|profiler with support for detecting cache misses. [..] |
|[..] |
| |
|[..] |
| |
|Please feel free to ask for clarifications. As I mentioned above |
|though, I may be too busy this week. |
| |
|Yours sincerely, |
|Colin Paul Gloster, someone with a computer science degree though |
|currently a poor Ph.D. student based at |
|Department of Physics, |
|Universidade de Coimbra, |
|Rua Larga, |
|3004-516 Coimbra, |
|Portugal. |
|Telephone: 00 351 239410663 |
|Fax: 00 351 239829158" |
|----------------------------------------------------------------------------------|

You mentioned hard disks. Before buying a new computer, you could try
running FLUKA with different hard disks on an old computer. However,
there are also different interfaces for hard disks. So, if you are
lucky enough to have computers which are similar except for the disk
interfaces (which is not very likely), then you could also run FLUKA
on those.

Yours sincerely,
Colin Paul Gloster

Thomas Cybulski <tomasz.cybulski_at_Quasar-Group.org> sent on August 1st, 2011:
|------------------------------------------------------------------------|
|"Hello! |
| |
| I am going to buy a workstation that will be mostly dedicated for |
|FLUKA runs. I am just wondering if anyone could give me some hints about|
|the optimum computer requirements that a FLUKA computer should have? |
|Particularly, I have in mind the hard disks, processor(s') speed and |
|#cores, RAM etc. |
| |
|Thank you very much in advance for your help, |
|Tomasz |
| |
|-------------------------------------- |
|Tomasz Cybulski |
|QUASAR GROUP |
|Marie Curie Fellow |
| |
|The Cockcroft Institute |
|4 Keckwick Lane |
|Daresbury |
|Warrington |
|WA4 4AD |
| |
|Department of Physics |
|The University of Liverpool |
|Liverpool |
|L69 7ZE |
|United Kingdom |
| |
|phone: +44 1925 864057 |
|e-mail: t.cybulski_at_liv.ac.uk |
|------------------------------------" |
|------------------------------------------------------------------------|
Received on Sat Aug 06 2011 - 13:39:14 CEST

This archive was generated by hypermail 2.2.0 : Sat Aug 06 2011 - 13:39:15 CEST