Implementing Shared Memory on Mesh-Connected Computers and on the Fat-Tree

AbstractWe present deterministic upper and lower bounds on the slowdown required to simulate an (n, m)-PRAM on a variety of networks. The upper bounds are based on a novel scheme that exploits the splitting and combining of messages. This scheme can be implemented on an n-node d-dimensional mesh (for constant d) and on an n-leaf pruned butterfly and attains the smallest worst-case slowdown to date for such interconnections, namely, O(n1/d(log(m/n))1-1/d) for the d-dimensional mesh (with constant d) and O(nlog(m/n)) for the pruned butterfly. In fact, the simulation on the pruned butterfly is the first PRAM simulation scheme on an area-universal network. Finally, we prove restricted and unrestricted lower bounds on the slowdown of any deterministic PRAM simulation on an arbitrary network, formulated in terms of the bandwidth properties of the interconnection as expressed by its decomposition tree