Abstract and subjects
If you were intent on building the world’s fastest supercomputer, how would
you design the interconnection network? As with any architectural endeavor,
there are a suite of trade-offs and design decisions that need to be considered:Maximize performance Ideally, every application should run fast, but
would it be acceptable for a smaller set of “important” applications
or application classes to run fast? If so, is the overall performance of
those applications dominated by communication performance? Do those
applications use a known communication pattern that the network can
optimize?