SC19 Proceedings

The International Conference for High Performance Computing, Networking, Storage, and Analysis

Poster 25: Leveraging Smart Data Transfer and Metadirective in Adaptive Computing


Student: Anjia Wang (University of North Carolina, Charlotte; Lawrence Livermore National Laboratory)
Supervisor: Yonghong Yan (University of South Carolina)

Abstract: In this work, we propose smart data transfer (SDT) along with extensions to metadirective and map constructs in OpenMP 5.0 to improve adaptive computing. The Smith-Waterman algorithm is used as an example, whose naïve implementation does not conduct data transfer efficiently. SDT is used to solve this issue with the following advantages: (1) SDT only transfers necessary data to GPU instead of the whole data set, resulting in 4.5x of speedup in our initial experiments. (2) Depending on computing vs. data transfer requirements of a program, SDT will transfer the output of each iteration from GPU to host either immediately or all together after the last GPU kernel call. (3) It supports large data exceeding GPU device memory's size via data tiling. We propose to extend metadirective's context selector to obtain similar improvement by enabling target enter/exit data and on-demand data access.

ACM-SRC Semi-Finalist: no

Poster: PDF
Poster Summary: PDF


Back to Poster Archive Listing