Does the Sharing of Execution Units Improve Performance/Power of Multicores?

Several studies and recent real world designs have promoted sharing of underutilized resources between cores in a multicore processor to achieve better performance/power. It has been argued that when utilization of such resources is low, sharing has negligible impact on performance, while offering considerable area and power benefits. In this paper we investigate the performance and performance/Watt implications of sharing large and underutilized resources between pairs of cores in a multicore. We first study sharing of the entire floating-point datapath (including reservation stations and execution units) by two cores, similar to AMD’s Bulldozer. We find that while this architecture results in power savings, for certain workload combinations, it also results in significant performance loss of about 28%. Next, we study an alternative sharing architecture where only the floating-point execution units are shared, while the individual cores retain their reservation stations. This reduces the performance loss to 14%. We then extend the study to include sharing of other large execution units that are used infrequently, namely the integer multiply and divide units. Subsequently, we analyze the impact of sharing hardware resources in Simultaneously Multi-Threaded (SMT) processors where multiple threads run concurrently on the same core. It is observed that sharing improves performance/Watt at a negligible performance cost only if the shared units have high throughput. Sharing low throughput units reduces both performance and performance/Watt. To increas