The author developed an open-source library to evaluate the performance of seven different Linux network backends, including DPDK, AF_XDP, and io_uring. The research focuses on measuring round-trip time (RTT) for packet processing across various configurations.
Testing was conducted on two distinct environments: a local desktop equipped with an Intel 82599 10GbE network card and a pair of AWS c6in.4xlarge instances. The benchmarks utilized a fixed-frequency load generator sending 100-byte messages at rates ranging from 20k to 1.6M messages per second.
The study highlights that interrupt handling and softirq processing significantly impact latency, particularly when competing for CPU resources. For instance, moving IRQ processing to a hyper-thread neighbor added 3 microseconds to UDP and 5 microseconds to io_uring performance on the Intel hardware.
The findings emphasize that backend selection is critical for performance tuning. While DPDK operates without interrupts, kernel-based solutions like AF_XDP and standard sockets rely on different mechanisms, such as XDP hooks or busy polling, to manage packet flow and memory access.
