Slashdot Mirror
First 96-Node Desktop Cluster Ships
Panaphonix writes "The Register reports that Orion Multisystems is shipping the first 96-node desktop cluster. 'With the new, larger system, customers get pretty much the most powerful computer around that can plug into a standard electrical socket.' According to the spec sheet, the DS-96 runs Fedora Core 2 and gets 110 GFlops sustained, 230 GFlops peak."
finite bandwidth between processors makes it impossible to sustain anywhere near peak performance for most real-world applications.
Linpack is what is usually used to measure sustained performance on HPC systems.
Nothing to see here; Move along.
The theoretical max gives a rough estimate of the raw Floating-point power for all of the processors on the system. You pretty much add up the GFlops potental for each node (not exactly, but pretty much). The sustained and demonstrated GFlops of the cluster is based on the Linpack benchmark. The reason there is such a huge difference between the two numbers can be a result of a few factors. 1) The interconnect is GigE and Linpack tends to make use of Message passing comms which are affected adversely by the latencies produced by the GigE connections (myrinet would have been a good choice, but I suppose it was probably impossible to squeeze that into that case) 2) Memory speeds also are a factor as pushing floating point numbers around involves memory. This cluster isn't using anything fancy when it comes to the memory and I suspect this may be another cause for this.
When they say that this line of clusters can "make or break" Orion, I am right now, leaning for broke. For the cost of this machine, one can get a real cluster with a lot more performance. I know this thing is nice because of the power requirements and the fact that you don't need a dedicated server room to store it, but for $100,000, you can get Microway to build you a pimptacular cluster with Dual-Opteron nodes, high-speed memory and a phat interconnect with either myrinet or infiniband. You will get a lot more work done for the same price.
Will this make or break Transmeta? It uses their processors (Transmeta Tinside as the Register calls it). Slashdot already pronounced the death of Transmeta though (it has no more niches!), maybe this could revive interest?
Without a proper flamewar, Anonymous was undecided on what shell to run.
Parent poster messed up on their calcs. Current XServes are 18.4 GFlops peak, not 35 eg Virginia Tech currently at #7/500 is 20240 GFlops peak for 1100 XServes. So 7 would only be ~129 GFlops peak, and 33 would be 607 GFlops peak. But not exactly fitting in a single tower case - though 8 would fit nicely one of those mini sound-padded racks which would be almost as good. And at least the last time I saw a price comparison made, the G5s were far cheaper than comparable rack P4s. (The G5 has 2x the FP hardware).