Performance of IP-forwarding of Linux hosts with multiple network interfaces

This paper evaluates and compares the performance of IP-packet forwarding of a Linux host equipped with multiple network interface cards (NICs), namely two receiving NICs and one transmitting NIC. We consider a Linux host with SMP (Symmetric Multiprocessing) or multicore multiprocessor (MCMP) architecture. We measure IP forwarding by subjecting an MCMP Linux host to different traffic load conditions of up to 1 Gbps. We used the IXIA hardware traffic generator to generate traffic with fixed- and variable-size packets. At the Linux host, generated packets are forwarded/routed from the two receiving NICs to the transmitting NIC. We consider two NIC affinity modes: (I) both receiving NICs are affinitized (or bound) to two cores of the same processor while the transmitting NIC is affinitized to a core on a separate processor, and (II) the transmitting NIC and one receiving NIC are affinitized to two cores of the same processor while the second receving NIC is affinitized to a core on a separate processor. For each affinity mode, we measure the performance for three packet reception mechanisms: NAPI (New API) with a default budget of 300, NAPI with a budget of 2, and Disable and Enable interrupt handling. The performance is measured and compared in terms of various key performance metrics which include throughput, packet loss, round-trip delay, interrupt rates, and CPU availability.