Memory Sharing on a CTA
•With
many processors referencing a single memory image, the probability that a given
reference
is nonlocal is (P-1)/P
•So, by
the CTA nearly all memory references will take l
time implying that each processor can do little work
•Responses
•A
“sea” of small processors that can be used “inefficiently”
•Have
many threads for processors to switch among
•Find a
way to greatly reduce l
•Change
processors from vN to something else