The paper, “To Move or Not to Move? Page Migration for Irregular Applications in Oversubscribed GPU Memory Systems with DynaMap,” was authored by Anand Sivasubramaniam, distinguished professor of computer science and engineering, and Chia-Hao Chang and Adithya Kumar, both computer science and engineering graduate students advised by Sivasubramaniam. There were 63 papers submitted for this award, out of which 19 were accepted as finalists before this one was selected for the award.
The research focuses on the memory overuse problem of graphic processing units (GPU), which are electronic circuits that can accelerate computing workloads, help render images on devices, access memory in a computing system quickly and more. When data is moved to and from the GPU continuously, as is the case when running large applications that access the memory irregularly, it can result in poor performance.
“Our detailed analysis of these executions reveals a very novel insight: rather than moving all data to the GPU, it is better to place data either on the central processing unit or GPU based on the spatial utilization,” Chang said. “Therefore, we built an adaptive data migration scheme, called DynaMap, to track data spatial usage at runtime and move data to the GPU accordingly.”
DynaMap enables an increase of more than two and a half times the speed of current state-of-the-art implementations, according to Chang.
The paper and its authors were recognized during the International Systems and Storage Conference, which was held June 14-16 virtually.
College of Engineering Media Relations
Last Updated July 08, 2021