Abstract
Although hardware support for parallel computation is increasingly available in heterogeneous embedded platforms, there is a distinct lack of effective system software support. One of the most significant challenges is communication between various computing elements for parallel implementation. Therefore, a data communication mechanism called streaming RPC was proposed to provide a system for modeling data streaming based on the client-server programming paradigm. As streaming RPC is based on the technique of buffering, the memory sub-system becomes a key factor affecting the facilities and performance. Because the number of streaming channels that a system can sustain are constrained by the architecture characteristics, this sets a limitation and raises difficulty when developing complex applications that need the amounts of channel over the bound to transmit data between program partitions. In this thesis, a concept of virtual streaming channels is proposed to increase the applicability and flexibility of streaming RPC. To eliminate the limitation caused by physcial resources, a block of memory called partition is shared by several virtual channels for transmition. The implementation of virtual channels mechanism is explained. The partition scheduler aims to schedule the arrived requirement of virtual channels with efficient policies respecting to response time requirements, and channel throughput. Therefore the dispatching strategis, latency-aware and throughput-aware scheduling is designed to satisfy different requirements. Furthermore, the partition scheduler applies different scheduling algorithms based on various concerns is also discussed. Finally, experimental results demonstrates the usage of the proposed methodology actually increases the flexibility of modified streaming RPC. Additionally, comparison results illustrating the differences between proposed scheduling policies and are given of each designed experiments. According to experiments, the latency-aware scheduling dispatchs the physcial partitoin with shortest response time. Moreovers, throughput-aware scheduling balances between latency and performance effectively.