Top 500 superComputers in the world site.

[Toaster]

Hi folks,
A site which shows the top 500 of the most powerful supercomputer sites in the world. Of course, these are Linux sites, Micky$oft is nowhere to be found.
One site, in antarctica, boasts 1500 nodes of rack mounted Compaq servers of modest configuration. These are "parallel processing arrays" meaning the complete node is dedicated to one task. It would take 35 Cray1 supercomputers to duplicate this amount of processor power. These systems do many things from Geological, weather forcasting and financial forcasting.
I hope you find this interesting.

http://www.top500.org

[Computer Hobbyist]

Very interesting site. Amazing how clusters of relatively low power processors have replaced the "big iron" of my youth. Could somebody explain the differences between clustered supercomputers and distributed computing as that term is used in the Seti and similar projects? More to the point are all such clusters parallel processing arrays simultaneously dedicated to the same task? Are they permanently dedicated or can they be easily reconfigured (read software) to perform multiple tasks as needed?

[Toaster]

In essence, these are called Beowulf arrays or parallel tasking arrays.
In almost all cases, the software or "task" must be parallel capable.
They can run any task so long as the above condition is met.
SETI@home is another matter. In this system, each system downloads a "work unit" or more like "a piece of the puzzle". Each system that uses the SETI client software does their own work unit the uploads the result(s) to the server to compile the data from other systems doing other work units. This is NOT parallel tasking/computing.
The BeoWulf arrays, like those in the supercomputing links are another matter. Each system is interconnected and runs THE SAME SOFTWARE or task. This works out to several (even thousands) of systems sharing CPU/Memory and other resources to do a common task. Often, one rather "beefy" machine does what I call "pool data" and does the final assembly to arrive at the end result.

The Cray-1 supercomputer of yesteryear was decomissioned some time ago in preference of the Cray-3 and Cray-4.
These supercomputers used "processor affinity" a method where each CPU in the system could be designated to a particular task.
Its a bit more involved then this but thats the basics.

[Computer Hobbyist]

I presume that all of the processors in a parallel tasking array must be located within a limited distance from each other and must have nearly identical hardware parameters? For example, is it correct to assume you can create a parallel array out of 1500 rack mounted Pentium III processors in the same building, but you can't array a bunch of mixed and matched processors spreadout in either time or space because of timing and integration issues?

_______________________________________________

Ron
Serious error.
All shortcuts have disappeared.
Screen. Mind. Both are blank.

[Toaster]

Hello again,
Actually, most parallel tasking systems are mix and match. While a configuration of alike hardware makes the parallel array easier to configure, it is not a requirement. It was common to use several 486 class machines to build an array.
Now, with low/mid level Pentium class systems being sold cheaply, most go this route. NASA used a rather large 486 array of I believe 1000+ systems not so long ago. Often, systems are now PII or P-Pro class systems. P-Pro systems do very well as the P-Pro is optimized for 32 bit code.
BeoWulf software optimization is now being done for the upcoming IA64 Intel CPU due out by the first of 2002. (in beta production now i`m told)