Load balancing in superscalar architectures

New techniques are increasing the degree of instruction-level parallelism exploited by processors. Recent superscalar implementations include multiple functional units, allowing the parallel execution of several instructions from the same application program. The trend towards an expansion of the number of hardware resources is likely to continue in future superscalar designs, and in order to maximize the processor throughput, the computational load must be balanced among these resources by the dynamic instruction-issuing algorithm. We investigate the effect on performance caused by the way instructions are distributed among the functional units of superscalar processors. Our results show that a performance gain of up to 38% can be obtained when the instructions are evenly distributed among the functional units.