Eager Sharing for Efficient Massive Parallelism

Workstation networks can become teraFLOPS supercomputers by adding highspeed interfaces supporting selective eager sharing. For Gaussian elimination and fast Fourier transform, selective eager sharing is much more efficient than global sharing of all data changes, and average efficiency remains above 60% for thousands of processors. Prototype SESAME interfaces will share data at 50 megabytes/second among more than 100 workstations. Propagation delays are typically 0:8 microseconds and overlap computations. All shared data reads are quick local accesses. Eager sharing supports diffuse nonlocal accesses in fine-grained parallel programs much more efficiently than demand driven cache protocols. Future massively parallel supercomputers should offer eagersharing coherence mechanisms. Key Words: Distributed Shared Memory, Write Consistency, Eager Sharing, Cache Update, Gaussian Elimination. 1 Introduction Multiprocessors, in which many processors equally share the same memories, limit th..