Of the two types of PE-to-PE communication available on the MasPar,
Xnet is the most efficient; Xnet can communicate data between
neighboring PEs at the rate of 5.75 GBytes/sec, whereas the global router
achieves only 325 Mbytes/sec (both figures are for a 4096 processor
machine). However, Xnet communications are only possible for local or
regular (same relative distance and direction) data communication; the
global router handles all irregular or nonlocal communication between PEs.
Xnet connects each PE to its corresponding neighbor PE; neighbors are indicated using the directions shown in Figure 1.
Figure 1: Xnet.
Consider the following PE array of 16 processors:
Suppose variables sum and summand reside on all the processors. If the statement
is executed, the values of summand on y and z will be added to sum on processor x. This is executed simultaneously on all processors of the PE array. Xnet treats processors on opposite edges as neighbors. For example, y would be considered the neighbor of v in the east direction, since the grid is logically arranged as if it were on the surface of a torus.