🚀 go-pugleaf

Since I care about parallel machines, my two cents worth:

FACT: we have proof-by-existence that

- massively parallel machines can be built, can be reliable, etc.
  (defining massive as "more than 1000 processor chips").
- they can have aggregate properties ( GIPS, GFLOPS, GB, GB/s IO)
  that are in the supercomputer league. Yes, I have details.
- they allow memory-intensive algorithms, since they can use, in main
  memory, the slower/cheaper DRAMs that Cray uses only in backing
  memory. Yes, I can back this up.
- for selected large applications, these machines already are the 
  fastest hardware, and the cheapest hardware. Yes, both.
- for selected applications, these machines aren't that hard to program.

IT SEEMS AGREED THAT:

- MIMD machines can have automatic load balancing, timesharing, etc.
  (Actually, the timesharing is called "spacesharing".)
- MIMD machines with virtual memory could conveniently fault pages 
  around between nodes.
- some applications could use a Connection Machine with millions
  of processors.
- programming isn't as easy as we'd like.

RESEARCHERS HAVE FOND HOPES THAT, SOME DAY,

- automatic parallelization onto these machines will become practical
  for many applications.
- shared-memory/cache-coherency will be cost-effective on large MIMDs.

MY OPINION:

- most supercomputer applications will wind up on massively
  parallel machines.
- there will aways be a market for the fastest possible single CPU.
- MIMD machines don't want the fastest possible node, because
  (so far) the money is better spent buying several cheaper nodes.
- conventional instruction set architectures are suitable bases for 
  MIMD nodes.
- "scaling laws" are sophistry when 8K node MIMDs are here now.

Sorry to be so wordy. 
-- 
Don		D.C.Lindsay 	Carnegie Mellon Computer Science