Improving the Performance of User-level Runtime Systems for Concurrent Applications

Concurrency is an essential part of many modern large-scale software systems. Applications must handle millions of simultaneous requests from millions of connected devices. Handling such a large number of concurrent requests requires runtime systems that efficiently man- age concurrency and communication among tasks in an application across multiple cores. Existing low-level programming techniques provide scalable solutions with low overhead, but require non-linear control flow. Alternative approaches to concurrent programming, such as Erlang and Go, support linear control flow by mapping multiple user-level execution entities across multiple kernel threads (M:N threading). However, these systems provide comprehensive execution environments that make it difficult to assess the performance impact of user-level runtimes in isolation. This thesis presents a nimble M:N user-level threading runtime that closes this con- ceptual gap and provides a software infrastructure to precisely study the performance impact of user-level threading. Multiple design alternatives are presented and evaluated for scheduling, I/O multiplexing, and synchronization components of the runtime. The performance of the runtime is evaluated in comparison to event-driven software, system- level threading, and other user-level threading runtimes. An experimental evaluation is conducted using benchmark programs, as well as the popular Memcached application. The user-level runtime supports high levels of concurrency without sacrificing application performance. In addition, the user-level scheduling problem is studied in the context of an existing actor runtime that maps multiple actors to multiple kernel-level threads. In particular, two locality-aware work-stealing schedulers are proposed and evaluated. It is shown that locality-aware scheduling can significantly improve the performance of a class of applications with a high level of concurrency. In general, the performance and resource utilization of large-scale concurrent applications depends on the level of concurrency that can be expressed by the programming model. This fundamental effect is studied by refining and customizing existing concurrency models