JPL Clusters XServes

burgburgburg writes "MacSlash has a brief note how NASA's JPL has put together a cluster of 33 XServes that was able to achieve 1/5 teraflop. The original article notes that the Applied Cluster Computing Group, using Pooch (Parallel OperatiOn and Control Heuristic Application) ran the AltiVec Fractal Carbon demo and achieved over 217 billion floating-point operations per second on this XServe cluster. More importantly, their research indicates that no evidence of an intrinsic limit to the size of a Macintosh-based cluster could be found."

Comments

Re:Where's the GigE switch?

[EricWright]

Maybe the computations were not communications bound... Fractal calculations can be done with a Monte Carlo method, which is highly parallelizable, and requires very little inter-node communication.

Re:No comparison?

[dhovis]

Apple provides libraries for doing double precision math with the Altivec unit. See here.

The theoretical peak performance for 33 XServes in the test done here was actually 495 GFLOPS, BTW. I don't know what the theoretical performance of double precision on Altivec is, though. LINPACK is all linear algebra (IIRC), so it would see some benefit.

I will admit that there are plenty of applications where the G4 is not the best processor available. I for one will certainly be happy to see the IBM PPC 970, but you shouldn't discount the XServe until the test is actually run.

Re:Where's the GigE switch?

[cremes]

The latency question is a good one. I'd say the answer to this lies in the driver for the NIC. I've written an IOKit ethernet driver and experienced pretty decent performance at 100 Mb. The system is processing packets as incoming data causes interrupts in the system.

However, I think the interrupt overhead for a 1000Mb link would be so high as to bring the machine to a screeching halt (okay, slow it down perceptibly). What a lot of driver writers do for gigabit links is to move their driver into polling mode. They essentially set a timer to go off every X milliseconds and process all the packets that have been copied into memory during that timeframe.

This gives a lower bound on the latency. A packet will always take X milliseconds to be noticed and processed by the system. Interrupt overhead stays low, but packet latency goes up a smidge.

It's a good trade off. I would bet that on a saturated link, packet latency at gigabit speeds is equivalent or WORSE than 100Mb. I might have to test that out...

cr

Re:Where's the GigE switch?

[Twirlip of the Mists]

I thought a lot of gig-e cards implemented a good deal of packet processing in hardware to deal with this very problem. Am I mistaken? I remember that the first PCI gig-e card I ever saw was installed in an SGI Origin, and when running full-out it pegged an entire CPU with interrupt handlers. Later versions of the card-- or perhaps an entirely different card, but sold by SGI and used with the same Origin servers-- had hardly any interrupt activity at all, even when moving data at rates exceeding 50 MB/s.