DATE: Thursday, October 24, 2024
TIME: 11:00 am - 12:00 pm
PLACE: NSH 3305
SPEAKER: Rishabh Iyer, UC Berkeley
TITLE: Performance Clarity for Systems Software and Hardware
ABSTRACT:
Performance is increasingly becoming a first-class requirement in system design. However, system developers today lack the techniques and tools to reason precisely about the expected performance behavior of both the code they write and the hardware they run their code on. Widely used tools such as profilers and simulators provide an incomplete understanding of performance, leading to hiccups and meltdowns in production when the workload or runtime environment changes in unpredicted ways.
In this talk, I will introduce a new approach to reasoning about the performance of systems software and hardware and make the case for designing new abstractions that precisely capture a system's expected performance behavior. I will present two tools that leverage these abstractions to help developers answer frequently-asked "what-if" performance questions. First, I will discuss CFAR, a tool that enables developers to reason precisely about how their code, as well as third-party code, uses the CPU cache. Then, I will introduce LTC, a tool that enables developers to reason about the potential performance benefits provided by hardware accelerators without requiring them to purchase the accelerator or port their code to it. The improved performance visibility provided by both CFAR and LTC has tangible benefits: for instance, we used CFAR to identify several cache-inefficient access patterns and performance bugs (including in the Linux kernel's TCP stack) and LTC to speed up compilation for ML accelerators by 5-41x.
BIO:
Rishabh Iyer is a postdoctoral researcher at UC Berkeley, working with Sylvia Ratnasamy and Scott Shenker. He received his bachelor's degree from IIT Bombay, and his PhD from EPFL, under the supervision of George Candea and Katerina Argyraki.
Rishabh's research is centered around developing techniques that enable developers to reason precisely about the performance behavior of their systems before they are deployed. His dissertation work introduced the notion of latency interfaces---simple, succinct programs that summarize a system's latency behavior just like semantic interfaces such as code documentation and specifications summarize functionality---and was awarded the ACM SIGOPS Dennis M. Ritchie Award, the Eurosys Roger Needham PhD Award, and the Dimitris N. Chorafas Award.
VISITOR HOST: Jan Hoffman
VISITOR COORDINATOR: Michael Stanley
SDI SEMINAR QUESTIONS?
Karen Lindenfelser, 86716, or visit www.pdl.cmu.edu/SDI/