Performance analysis and comparison of interrupt-handling schemes in gigabit networks

Interrupt processing can be a major bottleneck in the end-to-end performance of Gigabit networks. The performance of Gigabit network end hosts or servers can be severely degraded due to interrupt overhead caused by heavy incoming traffic. In particular, excessive latency and significant degradation in system throughput can be encountered. Also, user applications may livelock as the CPU power gets mostly consumed by interrupt handling and protocol processing. A number of interrupt-handling schemes has been proposed and employed to mitigate the interrupt overhead and improve OS performance. Among the most popular interrupt-handling schemes are normal interruption, polling, interrupt coalescing, and disabling and enabling of interrupts. In previous work, we presented a preliminary analytical study and models of normal interruption and interrupt coalescing. In this article, we extend our analysis and modeling to include polling and the scheme of interrupt disabling and enabling. For polling, we study both pure (or FreeBSD-style) polling and Linux NAPI polling. The performances for all these schemes are compared using both mathematical analysis and discrete-event simulation. The performance is studied in terms of three key performance indicators: throughput, system latency, and the residual CPU bandwidth available for user applications. As opposed to our previous work, we consider not only Poisson traffic, but also bursty traffic with empirical packet size distribution. Our analysis and simulation work gives insight into predicting the system performance and behavior when employing a certain interrupt-handling scheme. It is concluded that no single interrupt-handling scheme outperforms all other schemes under all traffic conditions. Based on obtained results, we propose and discuss a novel hybrid scheme of interrupt disabling-enabling and pure polling in order to attain peak performance under low and heavy traffic loads.