Compiling for shared-memory and message-passing computers

Many parallel languages presume a shared address space in which any portion of a computation can access any datum. Some parallel computers directly support this abstraction with hardware shared memory. Other computers provide distinct (per-processor) address spaces and communication mechanisms on which software can construct a shared address space. Since programmers have difficulty explicitly managing address spaces, there is considerable interest in compiler support for shared address spaces on the widely available message-passing computers. At first glance, it might appear that hardware-implemented shared memory is unquestionably a better base on which to implement a language. This paper argues, however, that compiler-implemented shared memory, despite its short-comings, has the potential to exploit more effectively the resources in a parallel computer. Hardware designers need to find mechanisms to combine the advantages of both approaches in a single system.