🚀 go-pugleaf

On 23 Dec 92 02:54:03 GMT, solman@athena.mit.edu (Jason W Solinsky) said:

solman> What I am trying to say is that when a task which has been
solman> non-cpu bound, is modified such that it becomes cpu-bound, the
solman> modifications invloved are typically excellent candidates for
solman> parallelization. Thus it still holds that "general purpose"
solman> computing can support the degree of parallelization available to
solman> a deep pipelined "hyperscalar" uP.

This is hopeful speculation.  Especially for the "deep pipelined" part.
Deep pipelines are nasty to exploit across basic blocks, not to say
across threads.

On the other hand, as you seem to argue, many current applications are
under-threaded, mostly because of the cost of threading under many
current OSes.

Still even many common applications can be broken in multiple threads
without much restructuring, and some people do it by hand; for example
recalculating a spreadsheet may be broken in as many threads are there
are independent clusters in the dependency graph between cells, and so
on recursively. Reformatting a document can be broken in as many threads
as there are independent sections, or manually imposed page boundaries.

I even remember that people at a french University (Poitiers?) had built
an Algol 68 compiler that used threads/futures to handle forward
references. I must confess that seeing that paper it had come into my
mind that one could just launch into a dataflow CPU a thread per each
syntactic phrase being parsed, and just let the CPU sort out all the
dependencies ensuring maximum parallel speedup, and I was conforted to
see Hewitt's parallel actors and garbage collection of processes ideas.

But then I started thinking of the likely cost/benefit ratio, and
repented.

The numbers I have seen so far seem to point out that on average
(some/many spreadsheets are poorly clustered, some/many documents have
no sections, in some/many compilations all statements depend critically
on the previous one...) the exploitable degree of parallelism is not
that great for most general purpose applications.

Of course one could claim that an as yet undiscovered architectural
solution will change everything; somebody once remarked, only half
jokingly, in this very newsgroup, that in twenty years time we will boot
ourselves in our massively parallel neural net engines, instead of
booting up our workstations, as we start our workday.
--
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