Re: more fun with proposals
(no name) (Marc Snir/Watson/IBM Research@nas.nasa.gov)
11 Jul 96 13:35:55
The alternative proposal is probably not more efficient than the original one.
It provides a higher level of abstraction than the original one, and is more
removed from the underlying h/w mechanisms. Thus, it may be less efficient for
some codes. The reason I have written it down and want it to be discussed is
that I am unconfortable with the large number of "auxiliary" functions and
concepts that we are introducing: counters, fences, coherence operations, etc.
I don't believe in magic: easier to use and more efficient. I want us to have
a discussion on the tradeoffs we are willing to make between ease of use and
efficiency on some machines.