Virtual memory for a hypercube multiprocessor

Most hypercube programs are structured so that all nodes contain an identical copy of the node program, as well as a complete copy of the node operating system program. This is a tremendous waste of memory, which ends up limiting the size and complexity of hypercube application programs. One way around this problem is to implement a virtual memory on the hypercube, whereby one copy of the node and operating system program is distributed among all the nodes of the hypercube and each node performs demand paging for the pages not resident in that node. Since almost none of the existing hypercubes have the hardware to support a virtual memory, this must be done in software. In this paper, we discuss a feasibility study of a hypercube virtual memory for executable code based on an implementation which requires no hardware support. We also explain die general principles involved in this type of virtual memory, and discuss how the features and restrictions of a hypercube architecture affect the implementation. In particular we show, via simulation results from real hypercube applications, ways to reduce die paging traffic in die hypercube