Overcoming scalability bottlenecks in Shenango

Abstract

In this thesis, I plan to present two new techniques that improve that scalability of Shenango, a recently developed research operating system that focuses on reconciling high CPU efficiency and high network performance for datacenter servers. Shenango relies on a centralized spinning component, the IOKernel, to deliver packets, monitor system-wide queueing delays, and effect core reallocations, all of which must complete every 5 microseconds. While this design is able to improve both CPU efficiency and networking performance over other systems, it limits the maximum packet-rate and maximum number of managed applications. I overcome these limitations with two new design elements: the first is an approach to flow-steering, DirectPath, that coordinates core-reallocation and packet delivery, removing the IOKernel entirely from the datapath. The second is a new kernel module called ksched that allows the IOKernel to reduces that amount of work spent on effecting core allocation by providing opportunities for batching and offloading work to remote cores.