A Retrospective on The MIT Alewife Machine: Architecture and Performance

The MIT Alewife project evolved out of exploratory work at Stanford on directory schemes for cache coherence [1] (also included in this issue). Using data from small bus-based multiprocessors, this early work demonstrated that directory schemes were as efficient as bus-based snooping protocols, and that by distributing directories along with main memory, they could provide the foundations for a cache-coherent shared-memory multiprocessor based on an interconnection network. This paper further recognized the scaling limits of bit-vector directories – they consumed memory proportional to the square of the number of processors – and speculated that variants such as limited pointer directories or limited broadcast directories might be attractive scalable alternatives. The paper, however, stopped short of demonstrating the feasibility of limited directories, largely because of the lack of either address traces or parallel programs written for a scalable coherent shared-memory system. This lack of data was not surprising given that such a machine had not been invented yet!