🚀 go-pugleaf

On 23 Dec 92 16:46:17 GMT, solman@athena.mit.edu (Jason W Solinsky) said:

solman> We seem to be saying the same things only you keep on talking
solman> about macro-level parallelism as though it would be difficult to
solman> exploit, and I feel that the existence of macro-level
solman> parallelism justifies increased superscalarity.

It's your very feeling that I find very, very optimistic, and totally
unsupported by past experience. But maybe I am too conservative, and I
hope that I am wrong. After all the human brain seems to be built like
you say, as a very highly parallel dataflow engine, not a massively
(macro) parallel machine.

solman> It is my contention that as processors continue to get more and
solman> more hyperscalar, they will be increasingly able to exploit
solman> macro-level parallelism without the help of smart compilers

As a contention, stated so bluntly, it stands on its own. I can only
repeat that I am skeptical, as the experience so far seems that basic
block level parallelism (which can exploited with multiple functional
units within the same CPU) is quite a different thing from continuation
level parallelism (which can be exploited with multiple CPUs within the
same machine).

solman> In a hyperscalar chip it will be necessary to implement tagging
solman> and/or scoreboarding protocols to control resource allocation
solman> and make sure that none of the data gets mixed up.

One of the big problems in going from the basic block to the
continuation level is that scheduling basic blocks is relatively easy,
as there are no conditional branches, or at most you deal with loops
(I am discounting heroic attempts to fit things like trace scheduling to
general purpose codes).

Getting multiple threads to run within the same CPU requires more than a
bit of things like speculative execution, and maybe even backtracking.

solman> Once this overhead is paid, the protocols can be altered to find
solman> macro level parallelism on their own with very little additional
solman> cost in either performance or area.

More than overhead I would say that the problem is that as of now nobody
have got the faintest clue on how to make this work in particular cases,
not to mention in the general case. I'd love to be proved underinformed
on this.

Maybe you are thinking of something like Hewitt's ideas on actor based
computing; but to me they look like an imaginative way organize macro
level parallelism.
--
Piercarlo Grandi, Dept of CS, PC/UW@Aberystwyth <pcg@aber.ac.uk>
  E l'italiano cantava, cantava. E le sue disperate invocazioni giunsero
  alle orecchie del suo divino protettore, il dio della barzelletta