New Supercomputer at NASA
from
RedNova
This will be interesting to follow since it's a 512-proc single system with NUMA (Non Uniform Memory Access) shared memory.
I worked with a group involved in scale-up of systems (cram more and more processors into a box) and it turned out that for most instances, you got much better performance clustering boxes together and splitting the computing into parcels than by trying to run a single instance of the program on one big box. You reached a point where the computer was spending more time trying to figure out what went where than working on the problem.
Cost is an issue too - Unisys makes a gorgeous 32-proc system which sells for about $2 mill. for a bare-bones system. You could buy four
very nice 8-proc boxes for about $400 K. Plus, if you had a total system failure with one of the boxes, you were still running at 75% capacity until it was repaired...
Posted by DaveH at November 18, 2003 12:32 PM