The structured memory access architecture: An implementation and performance-evaluation

The Structured Memory Access (SMS) architecture implementation presented in this thesis is formulated with the intention of alleviating two well-known inefficiencies that exist in current scalar computer architectures: address generation overhead and memory bandwidth utilization. Furthermore, the SMA architecture introduces an additional level of parallelism which is not present in current pipelined supercomputers, namely, overlapped execution of the access process and execute process on two distinct special-purpose, asynchronously-coupled processors. Each processor executes a separate instruction stream to perform its specific task which, together, are functionally equivalent in a conventional program. Our simulation results show that, for typical numerical programs, the access processor (MAP) is capable of achieving slip, i.e., running sufficiently ahead of the execute processor (CP) so that operand fetch requests for data items required by the CP are issued early enough and rapidly enough for the CP rarely to experience any memory access wait time. In this manner the SMA tolerates long memory access time, albeit high bandwidth, paths to memory without sacrificing performance. Speedups relative to the Cray-1 in scalar mode often exceed two, due to dual processing and reductions in memory wait time. 17 refs., 11 figs., 3 tabs