Ultra Fast Real Time Data Analytics through Heterogeneous CPU-GPU RDBMS
Project
Ultra Fast Real Time Data Analytics through Heterogeneous CPU-GPU RDBMS
Principal Investigator
University of Wisconsin-Madison
Oracle Fellowship Recipient
Oracle Principal Investigator
Weiwei Gong
Summary
This project aims to explore the design space of a heterogeneous CPU/GPU RDBMS and to build a database that fully exploits the performance benefits of GPU even when the data size goes beyond GPU memory capacity. The project explores four aspects of the system design: (1) Given a particular data placement, generate the best query execution plans for a workload so that both CPU and GPU are fully utilized. (2) Design the best data placement policy between CPU and GPU for high performance, (3) Design data compression algorithms in both CPU and GPU that strikes the right balance between compression ratio/speed, decompression speed, and recompression frequency. (4) Explore emerging architectural innovations including multi-GPU cluster, CXL, and GPUDirect.