Disk and Memory Recommendations

NOTE : For optimal performance, Intel’s quad-core “Nehalem” and six-core “Westmere” CPUs require memory configs comprised of DIMMS in multiples of 3.  In other words, optimal memory performance comes from configurations of 3GB, 6GB, 9GB, etc.  Therefore, for a Compute Node with 8 or 12 CPU cores, we recommend 12 x 1GB memory DIMMs.

Disk Striping

Both RTM and TOMO are I/O-bound applications, and disk stripping (aka RAID-0) is strongly recommended as an to maximize disk performance.  Hardware-based RAID-0 may offer better performance results, based on limited testing to date.  For an introduction to disk striping fundamentals, see the following :http://www.bestpricecomputers.co.uk/glossary/raid-0.htmhttp://www.webopedia.com/TERM/D/disk_striping.htmlhttp://searchstorage.techtarget.com/sDefinition/0,,sid5_gci214478,00.htmlhttp://www.hardwaresecrets.com/article/394

Operating System

Tsunami RTM is the only application in the current Tsunami Imaging Suite that requires 64bit Linux (Redhat 5.3 and gcc 4.1 or later are strongly suggested).  All other applications, including PSTM, PSDM, and TOMO are certified on both 32bit and 64bit versions of Linux, including : Red Hat, Suse, Debian, CentOS, and most others. 

More important than the OS version is the gcc compiler, which must be version 3.4 or higher.

Processors (CPUs)

All Tsunami applications are certified to run on most any server configured with Intel or AMD processors (x86 instruction set).  In 2009, Intel introduced a new CPU which we recommend strongly, as it delivered a major increase in performance (2-3X) over it's predecessor.  The CPU R&D project was code-named "Nehalem", and the marketing name is Xeon Series 55xx.  Here's a quick table listing the Intel processors most prevalent in the market, along with their official product series name, the associated year that it was introduced, and the unofficial code-name that's often used by IT professionals to refer to them:  

E53xx series : 2007 model. Code-Name "Clovertown"E54xx series : 2008 model  Code-Name "Harpertown"E55xx series : 2009 (this is what we have in Houston) Code-Name "Nehalem"E56xx series : 2010 (this is the first 6-core processor)  Code-Name "Westmere"

Processors (GPUs)

Tsunami RTM supports both CPU and Nvidia’s Tesla^TM GPU-based compute nodes.  New, integrated servers including CPUs and GPUs inside a single enclosure are being introduced continually; however, our application development, testing, and benchmarking is being performed on a cluster of nodes having the following configuration : One Nvidia S1070 (4 GPU cards, 4GB MEM ea, 2 shared x16 PCIe cables) connected to a Linux host server (dual quad-core CPUs, 12GB MEM, 4 x 250GB disks stripped, 2 PCIe x16 slots).
Here's a quick table listing the Nvida GPUs in the market today, along with their core count and a short summary of the breakthrough innovation that makes each generation special.  
"Tesla" is Nvidia's product family name for their GPU computing product line, including the latest introduction called "Fermi".  Nvidia has introduced 3 GPU computing products thus far, and our RTM application runs on all 3.  The product family now includes (including year of introduction, model name, core count, and major innovations) :

YEAR MODEL CORES INNOVATIONS
2006 Tesla 870 128 1.5GB MEM
2008 Tesla 1070 240 3GB MEM, More & Faster Cores
2010 Tesla 2070 448 64bit, ECC Memory

Networks

Standard Gig-E Ethernet is the best alternative for Kirchhoff applications, delivering performance scaling to 1000s of compute nodes with no discernible degradation in performance.  
TOMO doesn't rely on a network for performance or scaling.  Do keep in mind, though, that TOMO is I/O-bound, so the data files that it's working on should be housed on local disks rather than disks accessed over the network.
RTM is able to sustain peak computational performance when the shot that it is migrating can be retained in the node's physical memory.   When the shot size exceeds the size of the compute node's physical memory, RTM uses domain decomposition to spread the shot across multiple nodes.  When this occurs, a higher performance network such as Infiniband or 10-Gig Ethernet should provide substantial performance benefits.