Semantic-Preserving Transformations for Stream Program Orchestration on Multicore Architectures

Abstract

Because the demand for high performance with big data processing and distributed computing is increasing, the stream programming paradigm has been revisited for its abundance of parallelism in virtue of independent actors that communicate via data channels. The synchronous data-flow (SDF) programming model is frequently adopted with stream programming languages for its convenience to express stream programs as a set of nodes connected by data channels. Static data-rates of SDF programming model enable program transformations that greatly improve the performance of SDF programs on multicore architectures. The major application domain is for SDF programs are digital signal processing, audio, video, graphics kernels, networking, and security. This thesis makes the following three contributions that improve the performance of SDF programs: First, a new intermediate representation (IR) called LaminarIR is introduced. LaminarIR replaces FIFO queues with direct memory accesses to reduce the data communication overhead and explicates data dependencies between producer and consumer nodes. We provide transformations and their formal semantics to convert conventional, FIFO-queue based program representations to LaminarIR. Second, a compiler framework to perform sound and semantics-preserving program transformations from FIFO semantics to LaminarIR. We employ static program analysis to resolve token positions in FIFO queues and replace them by direct memory accesses. Third, a communication-cost-aware program orchestration method to establish a foundation of LaminarIR parallelization on multicore architectures. The LaminarIR framework, which consists of the aforementioned contributions together with the benchmarks that we used with the experimental evaluation, has been open-sourced to advocate further research on improving the performance of stream programming languages.