David Durst

Fourth Year CS PhD Candidate at Stanford
Advised by Kayvon Fatahalian and Pat Hanrahan

Photo of David Durst
I architect systems that make analyzing large data sets easier and more efficient. My current focus is simplifying the process of designing hardware accelerators for specific tasks including data-parallel pipelines and image processing. In the past, I've worked on decreasing the costs of using CNNs for computer vision tasks and using distributed computing to quality control and model large financial data sets. I'm supported by a NSF Graduate Research Fellowship and a Stanford Graduate Fellowship in Science and Engineering.


PLDI 2020

Designing efficient, application-specialized hardware accelerators requires assessing trade-offs between a hardware module's performance and resource requirements. To facilitate hardware design space exploration, we describe Aetherling, a system for automatically compiling data-parallel programs into statically scheduled, streaming hardware circuits. Aetherling contributes a space- and time-aware intermediate language featuring data-parallel operators that represent parallel or sequential hardware modules, and sequence data types that encode a module's throughput by specifying when sequence elements are produced or consumed. As a result, well-typed operator composition in the space-time language corresponds to connecting hardware modules via statically scheduled, streaming interfaces.

We provide rules for transforming programs written in a standard data-parallel language (that carries no information about hardware implementation) into equivalent space-time language programs. We then provide a scheduling algorithm that searches over the space of transformations to quickly generate area-efficient hardware designs that achieve a programmer-specified throughput. Using benchmarks from the image processing domain, we demonstrate that Aetherling enables rapid exploration of hardware designs with different throughput and area characteristics, and yields results that require 1.8-7.9x fewer FPGA slices than those of prior hardware generation systems.


Stanford AHA Monthly Meeting - October 2020
Reconfigurable accelerators promise an exciting set of benefits compared to other processors in the cloud and on mobile devices. They can enable application implementations that are more parallel, more energy efficient, and have lower latency. However, it can be challenging to predict the real-world situations where reconfigurability delivers these benefits. In this talk, I will examine five benchmarks that represent workloads interesting to Adobe. These benchmarks show that there is a precise niche of applications that benefit from reconfigurability: applications that can be implemented in a manner that takes advantage of custom cache hierarchies and specialized functional units. For other applications, there are other ways to improve performance including programming languages and compilers that efficiently use existing, non-reconfigurable accelerators with greater peak compute performance and memory bandwidth.

PLDI 2020 - June 2020
The conference talk for the Aetherling paper. In this talk, I focus on Aetherling's data-parallel IR with space- and time-types, a higher-level input language whose types are unaware of space and time, and a simple set of rewrite rules for converting from the higher-level language to the space-time IR.

Spark Summit East 2016 - February 2016
TopNotch is a framework for quality controlling big data through data quality metrics that scale up to large data sets, across schemas, and throughout large teams. TopNotch's SQL-based interface enables users across the technical spectrum to quality control data sets in their areas of expertise and understand data sets from other areas. I was the project lead and main developer for TopNotch while I worked at BlackRock.

Spark-NYC Meetup - September 2015
This presentation addresses the disparity between the current and desired big data user experiences. In this presentation, I demonstrate a web application with a scatterplot matrix visualization that allows non-technical users to utilize Spark to analyze large data sets.