Slashdot Mirror
Coolest Cluster Ever
sw155kn1f3 writes "Scientists at Los Alamos National Laboratory built a cheap (less than $1k per unit) 294-unit Beowulf claster dedicated to run astrophysics calculations. According to their website it's 85th fastest computer in the world. Seems cool and promising as it made with cheap components and off the shelf hardware."
It would be possible for a group of people, not necessarily a small group, but not necessarily huge, either to repeat this. 100 people, each with $3,000, could do it. The group would need to find some space to house the thing, and would probably have to do it in a climate where it could be relatively naturally cooled, which definitely rules out Phoenix. The computer would then be one of the fastest machines in the world.
Granted, I don't know what the hell they'd do with the computer, but it would be kind of cool to be on the list.
IBM had PL/1, with syntax worse than JOSS,
And everywhere the language went, it was a total loss...
Take a look at this room A 1000-Pentium Beowulf-Style Cluster Computer half way down the page.
Analytic & algebraic topology of locally Euclidean meterization of infinitely differentiable Riemmanian manifold
What's the bandwidth of that trunk? Also, what's the capaity of the connections between each 16-port card and the backplane?
Just curious... suppose all the units on a 16-port card have 1Gbps each, but only 8Gb total to the backplane. Then the backplane, in turn, has only 8Gb to the other switch. These are just made up numbers, but how would beowulf handle it? Can it group jobs requiring higher communication throughput onto the nodes which are closer to eah other? Does it have to be told the topology, or does it figure it out?
Sounds like they have 10 ports @ 1Gbit each free, so get about 10Gbit between the switches. When we (position 180 on the latest top500) were investigating thin-tree connections, we thought that we might be able to effectively run one job on a third of the nodes, another on the second third, and miscellaneous jobs on the 3rd (since we have two or 3 researchers who like to use lots of resources, and people like me who only need a single proc at a time). So you just partition the nodes and only allow your mpi jobs to sit on one group at a time.
Then there are things like openmosix which deal with topology automatically somehow. They will try to calculate the speed of the interconnect and the various nodes' procs (if heterogenous), and work out the distribution in the most efficient way. I am going to try to convince my sysadmin to try out openmosix on the lesser-used nodes of the cluster, because there is a feature of it that I think one of us might like - the combination of memory of the different nodes in one big contigous space, but right now we are busy cleaning up after the upgrade.