A Real Time MPEG-4 Parallel Encoder on Software Distributed Shared Memory Systems

摘要

This paper is dedicated to developing a real-time MEPG-4 parallel encoder on software distributed shared memory systems. Basically, the performance of a MPEG-4 parallel encoder implemented on distributed systems is mainly determined by the latency of data synchronization and disk I/O, and the cost of data computation. For reducing the impact of data synchronization latency, we invent a pipeline algorithm to minimize the number of data synchronization points necessary for video encoding. In addition, we employ a master-slave node structure to overlay computation and I/O in order for alleviating the impact of I/O latency. On the other hand, we propose a two-level partitioning method to minimize the cost of data computation, and overlap the encoding times of two different GOVs. We have implemented the proposed MPEG-4 encoder on a test bed called Teamster. The experimental results show the proposed MPEG-4 encoder has successfully met the requirement of real time through the support of previous techniques via 32 SMP machines, which are equipped with dual 1.5 GHz Itanium II processors per node and connected by Gigabit Ethernet.