Rishabh Iyer (UC Berkeley)- Performance Interfaces for Systems Software and Hardware

Abstract: Semantic interfaces—such as code documentation and specifications—provide simple, abstract descriptions of a system’s functionality, enabling engineers to reason about and use the system’s functionality without having to understand the implementation. In contrast, there exist no equivalent interfaces for system performance, despite performance having become a first-class citizen in system design. Popular approaches to thinking about performance, like benchmarking, profiling, and deriving performance envelopes, provide incomplete visibility, leading to frequent hiccups and meltdowns in production when the workload or runtime environment changes in unpredicted ways.

In this talk, I will introduce the notion of a performance interface and describe two techniques that derive simple, abstract performance interfaces for systems software and hardware, respectively. First, CFAR, which derives interfaces that enable precise reasoning about how systems code uses the CPU cache. Then LTC, which derives interfaces that enable engineers to reason about the performance of hardware accelerators. The improved performance visibility provided by CFAR and LTC has tangible benefits: for instance, we used the CFAR-derived interfaces to identify several cache-inefficient access patterns and performance bugs (including in the Linux kernel’s TCP stack) and the LTC-derived interfaces to speed up compilation for ML accelerators by 5-41x.