Optimizing energy-efficiency for multi-core packet processing systems in a compiler framework

摘要

Network applications become increasingly computation-intensive and the amount of traffic soars unprecedentedly nowadays. Multi-core and multi-threaded techniques are thus widely employed in packet processing system to meet the changing requirement. However, the processing power cannot be fully utilized without a suitable programming environment. The compilation procedure is decisive for the quality of the code. It can largely determine the overall system performance in terms of packet throughput, individual packet latency, core utilization and energy efficiency.ududThe thesis investigated compilation issues in networking domain first, particularly on energy consumption. And as a cornerstone for any compiler optimizations, a code analysis module for collecting program dependency is presented and incorporated into a compiler framework. With that dependency information, a strategy based on graph bi-partitioning and mapping is proposed to search for an optimal configuration in a parallel-pipeline fashion. The energy-aware extension is specifically effective in enhancing the energy-efficiency of the whole system. Finally, a generic evaluation framework for simulating the performance and energy consumption of a packet processing system is given. It accepts flexible architectural configuration and is capable of performingarbitrary code mapping. The simulation time is extremely short compared to full-fledged simulators. A set of our optimization results is gathered using the framework.ud