This issue requires some knowledge of the target machine, as far as processor layout is concerned. In order to simplify the situation, we assume that all communications between processors take the same time, in other words that no two processors are nearer each other than any others. This needs to be justified. On the Cray T3D, as compared to a distributed system such as a network of workstations, communications are relatively fast, so the differences between different sets of processors are small. In fact, the Cray has hardware support for a single address space, and this is used to implement the message passing mechanism that I will be using for the implementation. This implies that messages from a processor to itself are certainly quicker than from distinct processors, but that the time cost for other pairs of processors will be constant. So, we can still exploit locality, as we can allocate nodes to the same processor (no communication required) or to different processors (one unit of communication).