Achieving high CPU efficiency and low tail latency in datacenters

Abstract

As datacenters have proliferated over the last couple of decades and datacenter applications have grown increasingly complex, two competing goals have emerged for networks and servers in datacenters. On the one hand, applications demand low latency-on the order of microseconds-in order to respond quickly to user requests. On the other hand, datacenter operators require high CPU efficiency in order to reduce operating costs. Unfortunately, today's systems do a poor job of providing low latency and high CPU efficiency simultaneously. This dissertation presents Shenango, a system that improves CPU efficiency while preserving or improving tail latency relative to the state-of-the-art. Shenango establishes that systems today are unable to provide CPU efficiency and low latency simultaneously because they reallocate cores across applications too infrequently. It contributes an efficient algorithm for deciding when applications would benefit from additional cores as well as mechanisms to reallocate cores at microsecond granularity. Shenango's fast core reallocations enable it to match the tail latency of state-of-the-art kernel bypass network stacks while linearly trading throughput for latency-sensitive applications for throughput for batch applications as load varies over time. While Shenango enables high efficiency and low tail latency at endhosts, end-to-end application performance also depends on the behavior of the network. Thus this dissertation also describes Chimera, a proposal for how to build on Shenango to co-design congestion control with CPU scheduling, so that congestion control can optimize for end-to-end latency and efficiency.