Abstract: Models for Control Unit Synchronization on Shared Control Architectures

Shared control is a parallel architecture model that supports SIMD and MIMD operation using a centralized control configuration. Every PE supports a local instruction stream and receive its control by synchronizing with the shared control unit corresponding to its current instruction. Because the control is shared and not supplied on demand, the control units must be tightly synchronized to allow the PEs to receive control from the appropriate control unit for successive instructions in its instruction stream. This synchronization is accomplished by having the control units share a fundamental instruction cycle length; allowing the PEs to safely switch to a new control unit at the fundamental cycle edge. The fundamental cycle length is governed by the longest control unit length; all other control units are padded with NOP cycles to that length. The presence of very long instructions causes the fundamental cycle time to be excessively long. Since the fundamental cycle cost is suffered by each instruction, the performance of the system deteriorates. This paper discusses three solutions to minimize the effect of long control unit cycles on the fundamental cycle (and consequently the performance of the system). The implementation of the three techniques is discussed and the improvement in performance is characterized and contrasted with the implementation cost.
Back to Nael's homepage , or view my resume .