pc configuration, science, linux

Toggle sidebar Toggle sidebar
P

pnemeth

Honorable
Nov 6, 2013
8
0
10,510
Hi,

I am planning to get several computer nodes for science. These will be built from parts. The task is stellar model atmosphere calculations, that is, lots of matrix inversions. The executable code (fortran 77) is about 1.2 - 1.6 Mb, compiled with gcc 4.7, and the input data for the models is about 0.5 Gb.

I am looking for an optimal solution. I came up with this tentative configuration:

- CPU: AMD FX-8350 sAM3+ 8-core
- MB: Gigabyte GA-990-FXA-UD3
- Memory: Corsair XMS3 2x4GB DDR3 2000 MHz CL9
- SSD: A-Data SP900 Premier Pro 64GB Sata3
- CPU Cooler: Arctic Cooling Freezer A30

My question is: how could I tune this configuration to get the best performance of the system?

I plan to run Linux (openSUSE 12.3) base system on the nodes.

Thanks,
Peter
http://stelweb.asu.cas.cz/~nemeth/
 
Sort by date Sort by votes
COLGeek

COLGeek

Cybernaut
Moderator
Apr 6, 2009
54,136
3,391
179,340
Are these systems going to run 24/7? Also, can your code off-load processing to a GPU as well (CUDA or OpenCL)?

Do you also require HDD space?

If running 24/7, I would not overclock anything. You should consider a quality PSU as well (Seasonic, Corsair, XFX). Size (wattage) of the PSU will depend on whether you can take advantage of a GPU for processing your data (or not).
 
Upvote 0 Downvote
P

pnemeth

Honorable
Nov 6, 2013
8
0
10,510


Yes, they will run 24/7.



According to my understanding parallel processes benefit from GPU. The code
is sequential (iterative optimization) and a maximum speedup of ~3 can be
achieved with parallelization (OpenMP). But considering that the code
consists of ~250 subroutines and ~40 000 lines, I wish to avoid that
programming.
I want to calculate spectral models for a few 1000 stars, those can be done
independently, eg. parallel, but I cannot manage more than ~30 an a time.



No, not more than a few GB, what is a standard now. The process is rather
CPU intensive, so I think a fast CPU is the way with matching
motherboard+memory that can feed it.
Another question regarding the storage: What is faster nowadays a HDD or an
SSD? I chose A-Data because it supports ~500Mb/s r/w transfer rate.




Yes, a quality PSU is also important. However, as I don't think I can
use a GPU, I would like to save on that:

- Chieftec CTG-450-80P 450W

may suffice.

cheers,
Peter
 
Upvote 0 Downvote
COLGeek

COLGeek

Cybernaut
Moderator
Apr 6, 2009
54,136
3,391
179,340
A SSD is way faster than a HDD for loading the OS and any apps installed to SSD. This would also be true reading/writing data to/from the SSD vice the HDD. However, once the app is loaded and operating out of memory (assuming no swap), processing performance is the same (whether using a SSD or HDD).

In your case, I would have a couple more questions:

1. Your planned config calls for 8GB of memory. What do you expect your swap requirements to be? Swap on a SSD is much faster to access than on a HDD.
2. What GPU do you intend to use (unless you plan to remote access all of them)? Your planned motherboard has not on-board video. Seems you need something fairly basic.
 
Upvote 0 Downvote
P

pnemeth

Honorable
Nov 6, 2013
8
0
10,510


Then, SSD is the way to go here. Thanks.



8GB is good to start with, later it can be upgraded if necessary. The motherboard supports 32 GB, so I wish to devote 32+a bit GB swap space. The rest of the 64GB SSD is still enough for the OS and applications.



I plan remote access. I will get a single

- Sapphire HD5450 512MB DDR3 PCI-E

or similar, and after the system is up and running I wish to disable the GPU in BIOS. I did not check yet, but expect the motherboard allows it.

Peter
 
Upvote 0 Downvote
P

pnemeth

Honorable
Nov 6, 2013
8
0
10,510


I heard that the Intel compiler is generally faster by 20-30%, but I have never tried it. Until I can take a further step and invest in a fast compiler I would rather exploit the gcc/gfortran flags to tune it to the system and task. That's cheap.
I focus on the hardware at this stage, the compiler and OS can be reconsidered later. Also, I am not sure how an Intel compiler would perform with an AMD cpu. I guess this could fill an other thread here.

 
Upvote 0 Downvote
A

Alabalcho

Titan
Jan 13, 2011
14,019
435
76,290

Last time I have checked, Intel compiler had a free 30-days or so evaluation, so you can still benchmark your application before heavy investment in hardware (and this can even influence your hardware decisions). And I would think twice before putting SSD into number cruncher - you better spend these money on better CPU.
 
Upvote 0 Downvote
You must log in or register to reply here.

TRENDING THREADS

Latest posts

Moderators online

Share this page

COMPANY
RESOURCES
FOLLOW
Top Bottom