Parallelizing the GSM speech compression algorithm on Raw

Abstract

The GSM speech encoding algorithm, a widely used compression algorithm in mobile communications, was parallelized on Raw, a highly distributed hardware architecture with low latency communication. Parallelization architectures designed included a 2-stage pipeline and a 5-node satellite-hub model for distributing the time-intensive correlation calculations. Two additional architectures integrated these techniques: a 2-stage pipeline with a 4-node correlation parallelization, and an expanded 3-stage pipeline with a 9-node correlation parallelization. The expanded design took advantage of efficient switch-level data routing in the Raw network in order to distribute processing to non-local nodes. Results for the 3-stage, 9-node parallelization showed a factor of 3.6 speedup gain. A factor of 3.8 gain was achieved after optimizing data transfer. Future work includes extending the implementation to the entire GSM protocol, and utilizing available Raw resources to enhance the GSM speech compression algorithm with minimal performance loss.