On the performance of IP-forwarding for multicore multiprocessor Linux hosts

Abstract

Today's PC-based routers run mostly on Linux and have typical a mainboard with multiple processors and each processor has multiple cores. Such mainboard architecture is known as multicore multiprocessor (MCMP). Linux hosts with an MCMP mainboard can be configured for different network interface card (NIC) affinity modes and packet reception mechanisms. In this study, the authors study and compare the performance under these different modes and reception techniques for routing/forwarding IP packets. In particular, the authors consider two NIC affinity modes: (i) both NICs are affinitised (or bound) to separate cores of the same processor and (ii) both NICs are affinitised to separate cores on separate processors. For each affinity mode, the authors measure the performance for three packet reception mechanisms: NAPI (New API) with a default budget of 300, NAPI with a budget of two and disable–enable (DE) interrupt handling. The authors refer to these mechanisms as NAPI(300), NAPI(2) and DE. The performance is measured and compared in terms of various key performance metrics which include throughput, packet loss, delay, interrupt rates and CPU availability. The authors use the IXIA hardware-based traffic generator to generate traffic with fixed- and variable-size packets. At the Linux host, generated packets are forwarded/routed from one receiving network interface to a different transmitting network interface.