Towards portable end-to-end network performance characterization of SmartNICs
Towards portable end-to-end network performance characterization of SmartNICs
复制标题
DOI：
10.1145/3546591.3547528
复制
发表时间：
2022-08
期刊：
影响因子：
--
通讯作者：
Tong Xing;Hesam Tajbakhsh;I. Haque;Michio Honda;A. Barbalace
中科院分区：
文献类型：
--
作者：
Tong Xing;Hesam Tajbakhsh;I. Haque;Michio Honda;A. Barbalace
新兴的加速网络设备，包括配备ARM等通用CPU的SmartNIC(SNIC)，在硬件、网络堆栈和应用程序中产生了新的排队和处理延迟来源。因此，我们认为需要一个框架来测量基于SNIC的主机系统中的应用程序和网络堆栈性能。本文描述了我们的方法来构建这样的框架，重点是在主机或SNIC CPU上执行的常规Linux网络堆栈。我们的框架显示了SNIC如何影响端到端延迟、吞吐量和基于SNIC架构、SNIC CPU性能甚至应用本身的多核可扩展性。
Emerging accelerated network devices, including SmartNICs (SNICs) with general-purpose CPUs - such as ARM, create new sources of queuing and processing delay in the hardware, network stack, and application. We therefore argue for the need of a framework to measure the application and network stack performance in SNIC-based host systems. This paper describes our methodology to build up such framework focusing on the regular Linux network stack executed on either host, or SNIC CPUs. Our framework shows how SNICs impact end-to-end latency, throughput, and multi-core scalability based on the SNIC architecture, SNIC CPU performance, and even the applications themselves.