Parallel Computing: Current and Future Issues of High End Computing

Speculative multithreading (SpMT) increases the performance by means of exploiting specula-tive thread-level parallelism. In this paper we describe the Mitosis framework, which is a com-bined hardware-software approach to finding and exploiting speculative thread-level parallelism,even in the presence of frequent dependences between threads. The approach is based on predict-ing/computing thread input values via software, through a piece of code that is added at the begin-ning of each thread, the pre-computation slice (p-slice). A p-slice is expected to compute the correctthread input values most of the time, but not necessarily always. Because of that, aggressive opti-mization techniques can be applied to the slice to make it very short. In this paper, we also describethe microarchitecture that supports this execution model. The main novelty of the microarchitectureis the organization of the register file and the cache memory in order to support multiple versions ofeach variable and allow for roll-back in case of misspeculation. We also describe a novel compilerapproach to identify points where speculative threads are most effectively spawned, and generatethe corresponding p-slices for each. We show that the Mitosis microarchitecture-compiler achievesvery important speedups for applications that the compiler cannot parallelize by conventional non-speculative approaches, such as the Olden benchmarks.