Supporting a dynamic SPMD in a multi-threaded architecture

The authors present a multithreaded architecture model which can efficiently support a single-program multiple-data (SPMD) computation of programs with dynamic data structures. It is based on a dynamic SPMD model where the access delay due to a remote reference of a dynamic data structure can be tolerated by having multiple threads of control concurrently in execution within each processor. However, the present model permits the exploitation of locality of references through the use of caches for remote memory operations. When a remote memory access operation is encountered and cannot be satisfied locally, the processor can have the flexibility of migrating the thread to a remote processor when (and only when) such migration is desirable.<>