Slashdot Mirror
Blue Gene/L Tops Its Own Supercomputer Record
DIY News writes "Lawrence Livermore National Laboratory and IBM unveiled the Blue Gene/L supercomputer Thursday and announced it's broken its own record again for the world's fastest supercomputer. The 65,536-processor machine can sustain 280.6 teraflops. That's the top end of the range IBM forecast and more than twice the previous Blue Gene/L record of 136.8 teraflops, set when only half the machine was installed."
..figure out what the hell we are going to be doing for energy in 15 years??
"Look to the future and the present will be safe"
The damn thing's smarter than I am. Well, that's taking an estimate of 100 teraflops for the human brain, which seems to be popular.
Real_men_don't_need_spacebars.
Because it's at IBM, it's possible that it'll be used for their World Community Grid at some point, probably when it's idle... The current project there is, in fact, protein folding.
My Video Blog!
The legitimate thing that I can imagine is if it was a cost based contract that was given out before the cost of the hardware was known.
Was it?
I think this is more appropriate.
How we know is more important than what we know.
Back when It was only half installed I got to take a tour of it while it was in Rochester, MN... Got to walk through it and touch it. Turns out the computer that controls blue gene takes up about half as much space as blue gene itself.
Actually, the system is provided with the IBM XL family of compilers.
Notice that the performance has actually increased PER proccessor as you add more proccessors... This is very remarkable in computer technology.
Normally when you add cpus to a computer you get a increase in performance, but it doesn't increase linearly with each cpu. You have one cpu you have 100% performance, add one more and you may have 180% the performance and add 2 more you may have 300% of the performance etc etc.
Notice that with half the machine there it got 138 GFlops.
So if you doubled the size of the machine you'd expect to get something like 260 Gflops per second.
But you have 280 Gflops per second.
This pretty much means that as you add cpus the performance of each cpu actually increases slightly. That's a exponentional growth rate, at the beginning of the curve.
Of course there has to be a technical limit to the system and the amount of space, heat, and electricity it can handle.. but technically if you double the size of the cluster again I wouldn't be suprised if you'd get close to 750 GFlops per second performance.
This is some seriously hardcore stuff, the future of computing hardware. Todays supercomputer, tomorrow's desktop.. I can't wait.
65536 processors = 64K processors.
damn that IBM, they take geekiness to just a whole different place.
Funtage Factor: Purple
What do you think this machine is doing? You use this machine to calculate how far you can push your reactors to provide the energy you're talking about.
"There is nothing nice about Steve Jobs and nothing evil about Bill Gates." - Chuck Peddle
The main drawback to forecasting models is that it takes soo long to run all the data, so we have to cut back on the data so that we can actually see what's forecast before it happens. With this this thing running an expanded version of the GFS with 10KM resolution, we might be able to actually get it right for once. ;)
Someone save me from this sanity.
In theory, OpenMP allows you to do some nice things you can't accomplish well in MPI - load balancing the work being done on several CPUs on a node being one of the most important. However, in practice, the addition of OpenMP often complicates a code for little benefit because:
a) If your code load balances on a node but not between nodes, you're still rate-limited in domain decomposition type of problems by the data taking the longest step. There are strategies to minimize this, but when we're talking about codes running on 64K processors, this is already a big worry!
b) More importantly, the early benefit OpenMP gave over MPI when running on shared memory is no longer there! Most modern MPI distributions can optimise communication from one CPU to another (via MPI calls) that happen to be on the same node to use shared memory segments instead of sending the message out to the wire.
The added complexity of hybrid models are rarely worth it in these days of (generally speaking) large numbers of small-way SMP nodes. Perhaps systems like NASA's 20 x 512-CPU system are more amenable to such things, and with the upcoming Horus chipset for up to 32-way Opterons, this could see a revival, but MPI is by far the dominant paradigm, and rightfully so.