Locality-aware predictive scheduling of network processors

Demands for flexible processing have moved generalpurpose processing into the data path of networks. Processor schedulers have a great impact on the performance of these real-time systems. We present measurements that show that the workload of a network processor is highly regular and predictable. Processing time predictions, based on these measurements, can be used in scheduling together with information about locality in the instruction stream to significantly improve throughput performance. We propose two scheduling schemes, Locality-Aware and Locality-Aware Predictive, that try to avoid cold caches when scheduling packets for processors. Simulations of the schedulers using packet processing times obtained from an operational network processor show the tradeoffs between the algorithms and their performance improvements over First-Come-FirstServe scheduling.