Parallel Computer Architecture

11.6 Precommunication in a Shared Address Space 887

may not be prefetching early enough when memory latencies are

high.

Since we

prefetch all the children at one time, in most cases the depth-first work done for

the first child (or two) should be enough to hide the latency of the rest of the

children,

but this may not be the case. The only way to improve this is to

speculatively prefetch multiple levels down the tree when we encounter a

cell,

d

ere

f

erenc

¡

n

g speculatively prefetched pointers to determine prefetch addresses,

hoping that we will indeed touch all the cells we prefetch and they will still be in

the cache when we reach them. Sinc