Exploiting Speculative TLP in Recursive Programs by Dynamic Thread Prediction

Abstract

Abstract. Speculative parallelisation represents a promising solution to speed up sequential programs that are hard to parallelise otherwise. Prior research has focused mainly on parallelising loops. Recursive procedures, which are also frequently used in real-world applications, have attracted much less attention. Moreover, the parallel threads in prior work are statically predicted and spawned. In this paper, we introduce a new compiler technique, called Speculative Parallelisation of Recursive Procedures (SPRP), to exploit speculative TLP (thread-level parallelism) in recursive procedures. SPRP combines a dynamic thread-spawning policy and a live-in prediction mechanism in a single helper thread that executes a distilled version of a procedure on a dedicated core. It serves to predict both the invocation order of recursive calls and their live-ins in concert and dispatches these calls to the other cores in a multicore system for parallel execution. To our knowledge, SPRP is the first compiler technique to speculatively parallelise recursive procedures this way. Compared with existing static thread prediction techniques, dynamic thread prediction reduces the number of useless threads spawned, and consequently, misspeculation overhead incurred. Our preliminary results demonstrate that this technique can speedup certain recursive benchmarks that are difficult to parallelise otherwise.

A Static Greedy and Dynamic Adaptive Thread Spawning Approach for Loop-Level Parallelism

“…It is effective for speculative threads to reduce the cross-thread data dependences. Gao et al [24] utilized dynamic thread prediction scheme for predicting the invocation order of recursive calls and their relative live-ins to explore thread-level parallelism in the recursive programs. The hardware-based performance counters are used to predict the runtime performance profiles of speculative threads, which are considered as the metrics of the best loop level in the dynamic loop selection [15] .…”

CilkSpec

“…We employ speculation in an application-agnostic fashion to automatically identify such ready tasks. Speculation support for recursive programs designed by Gao et al [5] employs helper threads to execute past synchronization points but does not focus on the design of an efficient and accurate predictor.…”