Jonathan Mace - Research Statement

Distributed systems represent some of themost interesting and successful computing applications in use today, from web search and social networks, to data analytics and large-scalemachine learning, to loosely-coupled microservices, serverless lambdas, and the public cloud. However, it is notoriously diõcult to understand, troubleshoot, and enforce distributed systems behaviors, because unlike standalone programs they lack a central point of visibility and control: each component only performs a narrow slice ofwork, and to execute a high-level task (such as a search query or an analytics job) entails a complex ow across multiple processes,machines, and the network. When problems occur in a distributed system, their symptoms are o en far removed from the root cause, transient factors are o en to blame, and instead of an outright crash, systems experiencemore pernicious symptoms like sustained degraded performance. When they materialize in production systems (as they have with all major cloud providers) it leads to high-proûle outages and amassive loss of revenue. My research focuses on improving our ability to observe, reason about, and enforce end-to-end behaviors in distributed systems. _emain approach I have taken is to develop cross-cutting tools that co-ordinate across system components and layers at runtime, re-establishing end-to-end visibility and control over diòerent system behaviors. My work includes tools for co-ordinating scheduler parameters across shared systems to give users real-time performance guarantees (NSDI ’15) [9]; for monitoring metrics and properties to answer ad-hoc queries from users at runtime (SOSP ’15, Best PaperAward) [12]; and for recording structured performance traces and analyzing them in aggregate (SOSP ’17) [2]. A recurring theme ofmy work is to seek a global understanding of complex so ware stacks, peering through abstractions and layers to observe, and exploit, the end-to-end ow of execution. In practical terms this ranges from the underlying mechanisms for communicating information and orchestrating actions across component boundaries, to high-level abstractions that align the user’s perspective with the cross-cutting task, hiding system-level communication and coordination concerns. _roughoutmy research I have beenmotivated by real problems experienced bydevelopers andoperators in practice, and one ofmy primary goals is to design and implement practical tools that can feasibly be deployed in production systems. _is is especially delicate for cross-cutting tasks,which are diõcult to establish in practice because of ampliûed logistical costs – they require coherent developer choices and pervasive deployment across all inter-operating systems and components [11]. _is leads me to gravitate towards simple, elegant solutions to these otherwise messy problems, and I seek abstractions that minimize the eòort required to develop and deploy cross-cutting tools. I am also heavily involved with the distributed tracing workgroup, the main industry group for this topic, and I sit on the OpenTracing industrial advisory board. All ofmy research artifacts have been open-sourced; concepts introduced by my work have receivedmainstream adoption and are deployed in production at numerous companies. I will continue to collaborate with these groups, and other partners in industry, to identify challenges, validate solutions, and help advance the state-of-the-art In what follows I will present four areas ofmy current research, which paint a broad picture of the themes and overarching challenges in performing cross-cutting distributed systems tasks. Finally, I will outline future directions for my research that I intend to pursue in the next stage ofmy research career.