Publication
This thesis demonstrates that it is feasible for systems code to expose a latency interface that describes its latency and related side effects for all inputs, just like the code's semantic interface describes its functionality and related side effects.
Semantic interfaces, such as code documentation, header files, and specifications, are indispensable. By providing a succinct summary of a system's functionality, they make it possible for developers to efficiently reason about, use and deploy code they did not write themselves. In contrast, there is no equivalent construct that describes latency behavior in a way that is simultaneously succinct, precise, and complete. Widely-used representations such as envelopes (e.g., probabilistic upper bounds or asymptotic time complexity) or benchmarks (e.g., SPEC or TPC-C results) provide an incomplete understanding of latency, leading to hiccups and meltdowns in production when the workload or runtime environment changes in unpredicted ways.
We take a three-part approach to realize latency interfaces for systems code. First, we show how to design datacenter systems that provide predictable latency behavior while sustaining high throughput. We present Concord, an efficient runtime for datacenter applications that demonstrates how the careful approximation (as opposed to canonical implementation) of theoretically optimal scheduling policies enables datacenter systems to sustain significantly higher throughput while continuing to meet the same latency targets.
Second, we propose that the latency interface of a system be a program that accepts the same input(s) as the system and outputs its processing latency. We contribute three key ideas that help summarize latency in a succinct, precise, and complete manner: latency-critical variables, which provide succinct abstractions of how the system interacts with its environment, the latency resolution, which provides readers of the interface with explicit control over the trade-off between succinctness and precision, and deployment-specific interfaces which enable users of the system to reason precisely about its latency behavior in their distinct deployment environments. We concretize this representation in the domain of network functions (NFs) and present LINX, a program analysis tool that automatically extracts latency interfaces from NF implementations. We demonstrate that the \pix-extracted interfaces are succinct, precise, and complete and show how they can be used to identify latency regressions, diagnose and fix performance bugs, as well as identify the latency impact of NIC offloads.
Third, we present CFAR, a technique, and tool that allows developers to reason precisely about micro-architectural side effects (specifically CPU cache usage) of systems code. CFAR introduces memory distillates, an intermediate representation that contains all information relevant to how a program accesses memory and discards everything else. CFAR automatically extract