New! Oct 2019: We're presenting a system for checkpointing for DNNs to enable training enourmous DNNs on a single GPU.

Jan 2019: fiJIT is a high-performance system for DNN GPU multiplexing

Dec 2018: Revec: Program Rejuvenation through Revectorization was accepted to Compiler Construction 2019!

November 2018: Dynamic Space-Time Scheduling for GPU Inference was accepted to LearningSys (at NeurIPS 2018)!

Intro

I’m a Ph.D. student at UC Berkeley. As a Ph.D. student, I am advised by Ion Stoica and Joseph Gonzalez. I am a part of the RISE Lab, BAIR Lab and Berkeley Deep Drive Lab. I previously was a Research Scientist at DeepScale, a startup building tiny deep learning systems for autonomous cars.

Recent projects

Checkmate: Breaking the Memory Wall with Optimal Tensor Rematerialization
Paper link
The case for GPU multitenancy
Paper link
Revec: Program Rejuvenation through Revectorization
Accepted paper, Intl. Conference on Compiler Construction 2019
Paper link Github
Dynamic Space-Time Scheduling for GPU Inference
Accepted to LearningSys'18 at NeurIPS 2018
We demonstrate 2.5x to 4.5x speedups for GPU inference for deep learning workloads through a novel GPU multitenancy approach.
Paper link Github
DSCnet: Replicating Lidar Point Clouds with Deep Sensor Cloning
Latest paper from the DeepScale research team. We present a technique to replicate a $\$$70k LiDAR sensor with inexpensive sensors.
Paper link
Scalable Architecture for Anomaly Detection and Visualization in Power Generating Assets
Accepted to HPBDC'17 at IPDPS 2017
We present a scalable system for time-series anomaly detection that horizontally scales to sensor networks with >400k QPS.
Paper link Slides
Spotting Suspicious Reviews via (Quasi-)clique Extraction
Accepted poster at IEEE Security and Privacy 2015
We uncover suspicious activity by well-organized reviewer rings sponsored by Yelp. Our work sheds light on Yelp's little-known paid review operations.
Paper link

Research publications