Slashdot Mirror
Own a Piece of An Apple-Based Supercomputer
Graff writes "Now that Apple has come out with the Xserve G5, Virginia Tech has been swapping out parts of their 'System X' supercomputer for the more compact 1U Xserves. MacMall is selling some of those System X component G5 systems with an approximate $200 savings and an extra 512 megs of RAM over a normal G5. You can read more about it at MacCentral."
on the contrary. My 4 years experience with clusters show that after the first two months of burn-in where many components fail, you aftewards have a higher MTBF than with PCs used "normall", because in the Beowulf case the AC power is regulated, the machine is almost never switched on and off (major cause of damage because at startup every component consumes power at the same time, voltage drops, and damage occurs), temperature is kept constant, the machine is kept in a safe room where nobody ventures more often than once a week because of the cold and the noise, therefore there is no dust in the machines or grease on the contacts.
Google passes Turing test : see my journal
I'm going to have to disagree with you concerning the amount of wear and tear.
Most computers fail either in the first month or so of use or after many years of good use. In the first case it's usually a bad component that slipped by quality control. In the latter it is simply the ravages of time. Longer quality control "burn-in" times would eliminate many of those first month failures, but the vendor really doesn't have the time/space for long burn-ins.
Now, the Number One way to shorten the life of your computer is to turn it on and off frequently. The computer heats up when you run it and cools when it's off. The expansion and contraction of components associated with these temperature changes stresses every solder joint on every component -- and may even stress the chip-level components themselves. To lengthen the life of your hardware (at the cost of extra electricity), leave your system on unless you aren't going to be using it for a significant length of time (i.e.: don't power cycle more than once a day).
These G5's have been on for approximately six months straight in a very well-controlled temperature environment. This is a burn-in that virtually guarantees that there were no manufacturing defects. However, since they weren't power-cycling on a regular basis, it was actually a VERY low-stress environment.
Life is short: void the warranty.
You can do some good high speed networking using firewire.
It's available in Mac OS and Linux.
http://www.homenethelp.com/network/firewire.asp
400mbps isn't to be sneezed at. With repeaters it'd probably make a decent fail-over network in case the main gigabit link failed.
Worst
What I heard is that VT isn't removing the identification stickers. I don't know if MacMall is removing them or not.
First of all, how is this profiteering? They are trading in their G5s to upgrade their cluster, Virginia Tech is not selling the "old" ones, MACMALL IS! Virginia Tech is simply trading them in. So you would rather have Virginia Tech eat all the money that they spent to purchase the original lot of 1,100 instead of making them look much better to potential customers who want to purchase the use of their cluster, by leveraging the money they have already spent? It is in business to make a profit, like it or not, it is not their primary main objective (Chinpokomon!) but it can not be ignored, either. I respect your opinion that some of the G5s should have been kept around for use in labs, by students, etc. Then again, we don't know if VT traded ALL of their G5s, they just might have kept some for the uses that you and I laid out. I suppose this just shows, on a grand scale, the high resale value of Apples!
I hate sigs.
The G5 was the clear winner out of all the chips on the market, and Apple was the clear winner of the platforms considered, and they considered *ALL* of them worth considering.
The success of the venture simply proves the superiority of keeping an open mind and not bringing tired old pre-conceptions (Apple's slow, Apple sux, etc.) to your work.
Actually, FireWire supports transfers of up to 3.2Gbps depending on the interconnect used.
mbbac
Don't confuse PCI-X with PCI Express. PCI-X is PCI, clocked at 100/133 MHz. PCI Express is the former 3GIO technology
anyway if you look at the specs you can see all the silly stuff.... that cluster does not need 1100 Superdrives, or 1100 Radeon 9600 cards..... let alone size and whatnot... i'm sure it was done because the Xserves were just too far off and it was the only machine out there with the G5/970 chip for sale to anyone.
look at the specs:
The systems sold by MacMall are listed as 2.0GHz Power Mac G5s equipped with 1GB DDR SDRAM (2 512MB memory cards); equipped with 160GB ATA drives, a SuperDrive, ATI Radeon 9600 Pro graphics processor, Gigabit Ethernet, 3 USB 2.0 ports, 2 USB 1.1 ports, 2 FireWire 400 ports and 1 FireWire 800 port, along with an AirPort Extreme card slot and no modem -- in other words, a stock Power Mac G5 Dual 2GHz system with a memory upgrade from 512MB to 1GB
it does seem the pulled the fibre cards out... they are optional in Xserves... maybe they just swapped those? i don't know if they are the same in both machines normally.
I'm fairly sure Virginia Tech wanted the 1u cases all along (makes more sense). However, they needed the cluster up in time to make the Top 100 list. Being on that list brings in _lots_ of research money. So yes, they did need them.
Why?
Actually the list is updated every 6 months, but the one in November is the bigger of the two because that is the time for the annual supercomputing conference.
I believe refurb products have a 90 day warranty from Apple.
"massive wear and tear" is also known as "verified reliability"' to some people.
My sig is blank, I typed this by hand.
Nope. 3200 Mbits/s / 8 == 400 MByte/s.
1394b supports 800 Mbit/s over cat5, 1600 Mbit/s over poly-fiber, and 3200 Mbit/s over glass-fiber. Grandparent was right.
-T
That's an integer multiply-add. The PowerPC 970 can do a double-precision floating point multiple-add, and that is what the Opteron and P4 lack. They can get pretty decent throughput for this sort of thing using SSE2, but only about half of the throughput, clock for clock, that a PPC 970 can get.
Given that getting on the Top500 list seemed to the main goal of this system, and that list uses only the (very limited) Linpack benchmark which is essentially nothing but multiply-adds, this makes the PPC 970 a much better chip. Of course, for real-world code, the difference might not be nearly as large and in many situations the P4 or Opteron could easily be a lot faster.
Of course, one question that could easily come out of this is WHY doesn't SSE2 include a double-precision floating point multiply-add instruction? You would have to ask Intel about that one, because it seems like a natural instruction to have in SSE2 if you ask me. Even with the updated SSE3 they didn't add this.
Just called, no need to call.. they're all gone.. shucks.
-Jim
Not every. But most. Solving ODE's usually boils down to iterating a (possibly implicit) linear system. Solving PDE's with finite differences does too. Or with finite elements. Or spectral methods. Lots of statistical computations do too.
Certainly there *are* scientific applications that don't involve multiply-adds, it's just that the vast bulk of scientific computations that are suitable for parallelization really boil down to solving linear systems, some kind of linear iteration, least-squares problems, or some combination. All of which are solved using lots of multiply-adds. So, while linpack isn't the end-all and be-all of hpc benchmarks, i'd say that it's a pretty good guideline; i'd also say that the speed of multiply-adds matters a whole hell of a lot for scientific computing.