Stanford University Networking Seminar
A data centre hosts diverse applications and not all of them can be trusted, especially in the public cloud (e.g., Amazon AWS, Windows Azure). While virtualisation has made significant advances in isolating CPU performance, there's little to no protection for network bandwidth. Bandwidth contention between VMs can occur even if they do not communicate with each other. Moreover, this contention can manifest itself at timescales of a few milliseconds, invisible on human timescales, and we show how this can degrade long term performance.
In this talk I will outline one practical approach to virtualise network bandwidth and protect a tenant's network performance. Our system, EyeQ, works from the network “edge” (end-hosts) with minimal network requirements, with low CPU overhead at high line rates (10Gb/s), even with bursty UDP traffic. In our experiments, we found EyeQ protects a memcached cluster from a collocated bursty UDP tenant, bringing the 99.9th percentile response latency close to bare-metal performance.
We built EyeQ for Windows and Linux, and the Linux version is open-source: http://jvimal.github.com/eyeq. This is joint work with Mohammad Alizadeh, profs. David Mazieres and Balaji Prabhakar, and Changhoon Kim and Albert Greenberg from Windows Azure.
Vimalkumar Jeyakumar is a PhD candidate in the Computer Science department advised by Profs. David Mazieres and Balaji Prabhakar. He has worked on network performance virtualisation, verifiable network emulation and principled approaches to network debugging