Slashdot Mirror
NASA To Get 10,240 Node Itanium 2 Linux Cluster
starwindsurfer writes "US space agency Nasa is to get a massive supercomputing boost to help get its shuttle missions back in action after the 2003 shuttle disaster. Project Columbia, a collaboration with two technology giants, will mean Nasa's computing power will be ramped up by 10 times to do complex simulations."
10240x more dupes?
I want to delete my account but Slashdot doesn't allow it.
http://linux.slashdot.org/article.pl?sid=04/07/28/ 1427228&tid=163&tid=139&tid=103
Do the editors work for the USPTO as well?
// Agent Green (Ian / IU7 / KB1JQO)
// IEEE 802.3: All 10base Are Belong To Us
http://www.sgi.com/newsroom/press_releases/2004/ju ly/supercomputing_ctr.html
Rus
Cheap UK and US VPS
Think of it less of a win for Itanium and more of a win for SGI Altix (that happens to use Itanium). The SGI Altix machines have a single system image with 512 processors (there are 20 of these clustered together). As far as I know, this is actually the cheapest and highest performing system that can use 512 nodes in a single system image. Other choices (which I'm not even sure scale to 512 processors) include Sun (slow), Power (expensive), and MIPS (SGI predecessor to the Altix - slower). Also, they are working on methods to increase single system image size to 2048 nodes, I believe an industry first. Some workloads just like running in single system images much better than on clusters.
As for Itanium vs. Opteron - the Itanium kicks the Opteron's ass in floating point. Since NASA is presumably going to be doing a lot of engineering simulations, good FP performance is highly desirable. Having 6 MB of cache per node probably helps the Itanium beat out the Opteron for large memory footprint workloads as well.
Basically, until Cray releases Red Storm (not sure if they'll stay in business that long), an Opteron system doesn't exist that can match the performance of the SGI Altix.
Finally, Itaniums are NOT "rediculously more" compared to the 8xx Opteron line (which is the Itanium's real competitor in this area).
Dan
It does floating point a lot faster than Opteron.
It makes you believe this supercomputer is made out of commodity components.
That's blatantly false.
The SGI systems are highly proprietary equipments that provide very large bandwidth between the nodes, extremely low latency and tight integration. They're not regular Beowulf clusters. They really are single systems with hundreds or thousands of CPUs, all of them running the same single instance of the OS (as opposed to typical clusters which run one OS instance per node).
Because of the tight integration, the software does not have to obey the same constraints as when running on commodity clusters. Especially the requirement for total parallelization does not stand anymore.
Therefore, problems which cannot be translated into 100% parallel algorithms, and therefore do not run efficiently on commodity clusters, are easily tackled on SGI supercomputers.
That's why they can charge a high price on their systems - because they can solve problems that are not accessible to "normal" computers.
That being said, the system at NASA is indeed a cluster, but it's a "small" cluster (a handful of nodes), each node being a supercomputer with hundreds of CPUs. It's a hybrid that provides the best of both worlds.
Because it's a 20-nodes cluster, each node being a supercomputer with 512 CPUs.
u ly/supercomputing_ctr.html
The article was written, unfortunately, by a rather clueless journalist. Here's a link to the proper information:
http://www.sgi.com/newsroom/press_releases/2004/j
The "firmware" (the equivalent of BIOS) they have on the Altix is pretty damn smart, it's like an OS of it's own. It can do diagnostics, and inventory and a truckful of other things.
Powering up a huge complex beast such as an Altix is no easy task. You need lots of "intelligence" at the hardware level to do that.