Why is a locking scheme for four CPUs potentially not great for 64 CPUs?
Levy
@lilli
E.g. whether to use a test-test-and-set spinlock or a test-and-set spinlock, the first one may work better at 64 cores due to fewer communication; the second one may work better at 4 cores because of lower average latency
shpeefps
Decisions made in parallel computing often come to weight of trade-offs. Just like how earlier we used to consider thread overhead vs increase parallelism for speedup, here we are considering costs of communication vs latency to determine the best approach.
Why is a locking scheme for four CPUs potentially not great for 64 CPUs?
@lilli E.g. whether to use a test-test-and-set spinlock or a test-and-set spinlock, the first one may work better at 64 cores due to fewer communication; the second one may work better at 4 cores because of lower average latency
Decisions made in parallel computing often come to weight of trade-offs. Just like how earlier we used to consider thread overhead vs increase parallelism for speedup, here we are considering costs of communication vs latency to determine the best approach.