GTC 2020: QuickQuery: GPU-Based Approximate Query Processing for Sub-Second Exploration at Scale

GTC 2020 S21566
Presenters: Larry Rudolph ,Two Sigma Investments, LP; Steven Martin, Two Sigma Investments, LP
Abstract
Exploring data for signals can be viewed as a series of questions, queries, and hypotheses, each of which leads to either deeper exploration or a new line of inquiry. QuickQuery employs sophisticated sampling, exploits memory hierarchy, and leverages the power of GPUs to provide sub-second response times no matter the size of the dataset. Even with trillion-row datasets, researchers can execute many such queries at the speed of thought, all while providing confidence bounds on the results. QuickQuery lets the researcher specify constraints on both accuracy and the response time. It is highly tuned to be very performant; our proof-of-concept implementation on a single GPU server was more than 500 times faster than a 400-core Spark implementation for simple queries, and more than 1,000 times faster for more complex ones.

Watch this session
Join in the conversation below.