Performance Evaluation of a Hybrid Interrupt-Handling Scheme with Multiple Network Interfaces & SMP Configurations

Abstract

NAPI is the latest scheme used in the Linux networking subsystem to handle the reception of incoming packets. It has been introduced to mitigate the excessive overhead of interrupt processing associated with the incoming traffic especially of Gigabit Ethernet. However, NAPI still has some shortcomings and its performance can be further enhanced. In a previous research study, a hybrid scheme was introduced to enhance the performance of NAPI for user-level applications. The hybrid scheme combines the schemes of disabling and enabling the interrupts with NAPI. The hybrid scheme shows superior performance over NAPI for user-level applications. In the present research, we study experimentally the performance of the hybrid scheme for IP forwarding at the kernel and user levels. We consider different experimental setups including single and multiple network interfaces with uniprocessor and multiprocessor architectures. The Code for the Linux kernel and drivers is modified to support and implement such setups and configurations. The performance is evaluated experimentally in terms of latency, throughput, and CPU availability. IXIA hardware traffic generator/analyzer and open-source software traffic generators are used to perform measurements.