Hyfi: Architecture-independent parallelism on a cluster of multiprocessors. October 2002. [RTY + 87] Richard Rashid, Avadis Tevanian

A network of parallel workstations promises cost-effective parallel computing. This pa-per presents the HyFi (Hybrid Filaments) package, which can be used to create architecture-independent parallel programs—that is, programs that are portable and efficient across different parallel machines. HyFi integrates Shared Filaments (SF), which provides parallelism on shared-memory multiprocessors, and Distributed Filaments (DF), which extracts parallelism from net-works of uniprocessors. This enables parallelism on any architecture, including homogeneous networks of multiprocessors. HyFi uses fine-grain parallelism and implicit shared-variable communication to provide a uniform programming model. HyFi adopts the same basic execution model as SF and DF; this paper discusses the modifications necessary to develop the hybrid system. In particular, HyFi modifies the signal-thread model as well as the software distributed shared memory of DF. It also unifies the SF and DF reduction operations as well as the dynamic load-balancing mechanism of fork-join filaments. Application programs written using the HyFi API can run unchanged on any architecture. Performance is encouraging on fork/join applications, where excellent speedup is achieved. Also, the fine-grain model of HyFi allows up to a 14.5 % improvement due to overlap of communication and computation. Unfortunately, we find that iterative applications do not speed up well due to the inability of the Pentium Xeon architecture to efficiently support concurrent memory accesses.