An Alternative Fast Packet I/O with Native System Calls

Abstract

This paper proposes an alternative high-speed packet I/O that enables user-space applications to receive and transmit packets at high speeds using the read() and write() system call families. Recent high-speed packet I/O frameworks have opened up new application designs optimized for network performance. However, they have two obstacles to implementing applications: restraints on programming models and limitations on NICs. These frameworks provide applications with APIs to access packets in kernel space without memory copies instead of well-generalized APIs, such as the socket API. Thus, the applications would be tightly coupled with the framework-specific APIs, and would be forced to adopt the APIs' programming models. Moreover, typical frameworks require specific NICs, which the frameworks support for their optimization. In contrast, the proposed packet I/O, called hpio, adapts native system calls for the packet I/O API and works with native device drivers. These advantages offer developers an easy way to implement various high-speed applications. Our measurement results demonstrate that hpio achieved over 40 Mpps with 60-byte packets, although it involves memory copies on system calls and no optimization of packet buffers.