SimdFSM: An Adaptive Vectorization of Finite State Machines for Speculative Execution

Abstract

Parallel execution of a Finite State Machine (FSM) is challenging due to strong data dependency. Previous work proposed speculative execution to distribute the workload to multiple threads. While without dependent data, threads working from the middle of the input speculate multiple states, possibly resulting in redundant computations. Advanced efforts have achieved significant performance improvements in each thread using SIMD gather/shuffle instructions.

This paper studies various SIMD-based strategies in depth, with the following factors considered: (1) FSM size, (2) processor microarchitectures, and (3) SIMD instructions used. We present SimdFSM incorporating various methods and profile their performances using a real-world FSM collection under different configurations of these factors. The results show that the performance differences among these methods can be significant, and the winner varies depending on these factors. Thus, a wrong choice can result in unexpectedly poor performance.

Therefore, we design an adaptive strategy using the profiling data to select the best method among the ones available under the current execution environment. The adaptive strategy further samples states’ distribution in an input fragment to improve speculation success probability. The results show that it can always select the best method with an ignorable overhead.