1 A New Heterogeneous Packet Processing Architecture and Its Analytical Performance Model

Abstract — Today’s IP routers have to simultaneously meet multiple requirements such as programmability, scalability, power, and price besides the traditional objective of high throughput. Software routers like Click offer the best flexibility but suffer from a lower level of processing throughput. A few recent works prove the potential of Graphic Processing Units (GPUs) for high-speed packet processing. However, current GPU architectures cannot guarantee quality-of-service (QoS) of IP routing due to the batched execution model. In this work, we propose a novel heterogeneous, integrated CPU/GPU microarchitecture, Hermes, which adaptively maintains a balance between packet latency and overall throughput. A complete set of router applications are implemented on this architecture. Experimental results show that Hermes achieves a 5X enhancement in throughput, a 81.2% reduction in average packet latency, and a 72.9 % reduction in delay variance, when compared with a GPU accelerated software router. A byproduct of this research is an analytical model that catpures the throughput and processing latency of Hermes-alike heterogensous architecture. In this work, the model can be used to quickly estimate IP routing QoS metrics of different Hermes microarchitectural configurations under various traffic patterns. Simulation results reveal that the analytical model can accurate predict QoS metrics with an average error rate of less than 10%. Index Terms—Routers, SIMD processors, Heterogeneous (hybrid) systems.