Slashdot Mirror
NASA Installs Linux Supercomputer
unassimilatible writes: "Federal Computer Week reports that NASA plans to study the ocean's future with the help of the world's first supercomputer of its kind to run on the Linux operating system. The new supercomputer -- an SGI AltixT 3000 single-system image supercomputer -- has been installed at the space agency's Ames Research Center in California."
The SGI Altix 3000 is not quite a supercomputer. Our local university got the very first model for production use of the Altix 3000's successor, the Altix 3700, in last April or so, and it made it in the TOP 500 supercomputer list in last June, but it fell out of the current list. And the 3700 is even faster than the 3700, so what's so special about it?
A monkey is doing the real work for me.
this does not even make the top500 so by definition it is NOT A SUPERCOMPUTER.
top500.org does show that for 3.2 million of macs and under 2 million of infiniband and hardware racks you can get to the THIRD position in top500,org using macintohses and the mac os x.
The only IRIX technology they're grafting is the SCSI layer. With the 2.6 kernel, there really isn't that much they need to do. They're testing their 512 processor systems on linux 2.6 as we speak, btw.
Well, they pay lots of people for doing other stuff than space research. Just think of Larry Wall, who was working on Perl during his time at the NASA JPL.
A monkey is doing the real work for me.
So anyone got a copy of /proc/cpuinfo from this mother?
(Oh and do you do something special to 'top' so it doesn't give you 512 lines of CPU state?)
The speedup I quoted is per processor speedup, but AFTER they had done some tuning. After you just port but before the optimization it seems quite normal to get slower code rather than faster... but from what you said that is fairly common. I am not sure it should be so expected going from one NUMA MIPS machine to a NUMA Itanium2 machine. Apart from the fact that the Itaniums are not very well balanced..
I do think they will get it right eventually, but it is still bleeding edge stuff unfortunately.
Note this is what the computer people at the computer centre told me not my personal code. The current state of my personal code is that it runs at 1/4 of the per processor speed on the new computer, (i.e it is running about about 1/10 to 1/20 of the speed you would expect) and gives the wrong answer... Clearly those things could be related... but I am destinctly underwhelmed at the moment.
So... since it's "easy" to be done in hardware and "hard" to be done in software, hardware engineers are better at software optimization than software guys? ;)
No that is not quite what I meant. The best way has to be some combination of the two IMNSHO. In the Itanium they started with a blank piece of paper, and asked the hardware guys what they wanted. Like normal engineers they made the tough stuff (the optimisation) someone elses problem. All the itaniums do is do exactly what the compiler tells them to do. So the compiler has to work very very hard to get optimsed code, and you always have more information at runtime than you do at compile time.
What we are seeing from this is exactly what you would expect, something with a VERY good peak performance that performs well on some benchmarks, but royaly sucks for real world applications..
Buggy compilers are always a pain, but ignoring that, there's no reason why the compilers can't get better over time to produce better code. I know that doesn't help you right now though :)
I am sure they will get better... Intel are patching their compiler about every month, and have a huge number of people working on it. Whether the whole EPIC idea will work on the other hand I don't know.... They are going to struggle very hard to get anything like the peak numbers out of it for normal applications..