A Dual-Length Path-Based Predictor for Thread Prediction

In speculative multithreading (SpMT) architectures which exploit thread level parallelism from a sequential program, predicting and spawning threads that follow a correct control flow is a major performance factor. Strategies for this control speculation issue can be divided into two approaches. The first approach is to support a limited control speculation, in which threads may only be spawn at specific points in the program. For example: loop iterations, function call and returns, or control equivalent points. This approach minimizes thread misprediction by only speculating at highly predictable places. The second approach is a more aggressive one, in which threads are predicted and spawned as soon as possible. This latter approach bears a larger potential for performance improvement, given that the thread predictor is sufficiently accurate. This paper targets architectures that use the latter approach. It first studies the predictability of thread addresses using a dynamic path-based predictor. The results show that, in an alias free situation, a high prediction accuracy can be achieved by including sufficiently long path information. However, the requirements for prediction table entries are also increasing almost exponentially when longer path information included. For a finite size table, there are cases when a longer path, due to capacity aliasing, leads to a severely deteriorated accuracy. To overcome the problem, this paper introduces a dual-length path-based prediction technique. We combine two path-based predictors: one predictor is indexed using a shorter path information, while the other predictor is indexed using a longer path information. A selection table is added to dynamically select which predictor’s outcome to use. The rationale behind this hybrid technique is to have the shorter path predictor, which has lower capacity requirements, to redeem the situation when the longer path predictor suffers from a severe capacity aliasing. Furthermore, by manipulating predictor’s table update policy, threads that can be accurately predicted by one predictor can be filtered out from contaminating the other predictor. Simulation results show that the combined effect significantly improves the accuracy of thread prediction.