一种面向流媒体的多核处理器分组调度算法

Packet scheduling algorithm on multi-core network processor towards streaming media

针对流媒体分组处理和多核网络处理器cache亲和性的特点,提出了综合流调度和分组调度优点的两级调度算法,即FBLA。FCFS调度算法可以达到分组级的细粒度负载均衡,但cache亲和性却很差。基于hash的调度算法可以保证很好的cache亲和性,但难以保证核间负载均衡。FBLA算法对这两种算法进行了折中,既通过cache亲和性提高处理器利用率,又能够达到细粒度的核间负载均衡。理论分析和仿真评估表明,FBLA算法具有良好的cache亲和性和负载均衡性,转发延迟和延迟波动比FCFS算法更低。在亲和因子较小时,FBLA的平均转发延迟略高于基于hash的算法,但其他情况下都要优于基于hash的调度算法。

With respect to the characteristics of streaming media packet processing and cache affinity of multi-core network processor,this paper proposed an improved packet scheduling algorithm,called FBLA. While FCFS could manage fine grain load balance at packet level,it hardly benefited from cache affinity in flow-based packet processing. Hash-based scheduling algorithms always had good cache affinity,but they often caused load imbalance. FBLA managed to keep high processor utilization with good cache affinity,featuring the benefits of both FCFS and hash-based packet scheduling algorithms. Detailed analysis and simulation show that FBLA always performs better than FCFS. When the affinity factor is low,the mean value of FBLA is a little higher than hash-based algorithm. In other case,FBLA performs better than hash-based scheduling.