Colloquium Details
Efficient Data Communication Abstractions for Networked Applications
Speaker: Deepti Raghavan, Stanford University
Location: 60 Fifth Avenue Room 150
Date: March 15, 2024, 11 a.m.
Host: Anirudh Sivaraman
Synopsis:
Datacenter networks have become at least 20x faster in the last 15 years, while CPU performance has stagnated over the same time period. In public cloud and wider area networks, disaggregated storage is becoming more popular as dataset sizes continue to grow. However, application performance in these environments is still limited by outdated data communication abstractions. For example, Protobuf, one of the most popular serialization libraries, was designed 20 years ago, when Ethernet speeds were 10 Gbps and available compute cycles were not the bottleneck. My research focuses on evolving data communication abstractions for today’s environments. I demonstrate that using new techniques to allow application runtimes to optimize when, where, and how much data is moved across the stack leads to significant performance gains across three application domains.
With Cornflakes, I designed and implemented a new data serialization abstraction that offloads data movement from the CPU to the NIC, leading to 15-45% throughput gains over existing serialization libraries like Protobuf, without any new hardware. With Posh, I designed a shell runtime that automatically parallelizes and distributes unmodified shell scripts which access remote data, using a new abstraction called shell annotations, leading to 1.6 to 15x better latency. Finally, I will discuss ongoing efforts to design Alto, a network orchestration layer for applications that chain together language models and external tools, which achieves 2.8x higher throughput in a sample claim verification pipeline, by streaming data between the language model and downstream tasks.
Speaker Bio:
Deepti Raghavan is a final year Ph.D. candidate in the Computer Science Department at Stanford, advised by Philip Levis and Matei Zaharia. Her research interests are in systems and networking, particularly in designing and building systems with efficient communication abstractions to improve the performance of networked applications. She has received the Distinguished Artifact award at SOSP and an NSF Graduate Research Fellowship.
Notes:
In-person attendance only available to those with active NYU ID cards.