Originally posted by legion
I'm guessing the real rush here is that AMD has recently had a lot of major contracts to drop new clusters into the top 10 of the Super 500, which could knock Apple's cluster out of the top 10 before it even gets a chance. That would be a huge loss for Apple and it's marketing department (especially as a first attempt into the 500 which could spur more orders) I'm guessing IBM is also helping out since they have much more experience in these super-structured computers. Otherwise, we would have heard of a huge hiring spree in Apple's support center to manage such a cluster (as it stands now, it'd take the whole lot of them to support just the 1100 G5s at VT)
There are some caveats that need to be placed here:
IBM is the vendor on contract to build most of these
AMD Opteron clusters so it is never a case of Apple/IBM v. AMD or some such. Right now, IBM is the only tier-one vendor of AMD CPUs, so nearly all high scale HPC clusters will go through IBM. (Yes, you can build a renderfarm through
RackSaver, but this isn't the same thing.) Besides, AMD rents IBM Fishkill for design and fabrication.
To my knowledge there are only three Opteron clusters in the works that would be eligible for the Top500:
Japans AIST 2116CPU @ 11 Teraflops,
China's Dawning Information 2000 CPU @ 10 teraflops, and
Red Storm 10,000CPU @ 36Tflops. If such a Mac cluster were to be built, it would beat out the second and be completed before the third: this is hardly able to push such a computer from 3rd out of the top 10!
I agree that it doesn't seem practical to have a cluster up and running in time for the Fall
Top500. Usually clusters take a while to set up. This leads me to think that the 2200 G5 cluster of VATech only slightly related to the 2x2Ghz delays.
Actually Mac OS X makes a better base OS to cluster than Linux. Yes, more research has been done
clustering Linux boxes, but Mach is definitely superior in this regard. Besides, I'd imagine they wouldn't focus on handling clustering at the kernel level, but would run an application level message passing clustering system like MPich. I have no idea if LinPack can be run on top of this layer since I don't do HPC. Macs have been clustered on small scales for a long time this way.
As for waiting for XServes, I wonder what everyone is smoking. Clearly the thing to wait for is the 4-way or 2-way PPC970 IBM Blades and rackservers (1Q 2004) with AIX and the #1 vendor in the top 500 than a G5 Xserve with an uncertain release date and a vendor with no previous clustering experience.
Finally, in the PC world, the fan is controlled by the BIOS, not the OS. This is from experience, because I set machines to always on full fan when doing memory tests--in one case, my computer kept shutting down because the fan would keep turning off, thank god it was a Pentium.
I don't believe this is the reason for the delays, but it does allow one to be amusing. As one of my friends put it, ``Damn them, Virginia Tech probably has
my machine.''