Accelerating Data-Parallel Primitives and Multi-way Joins on Heterogeneous Processors

PhD Thesis Proposal Defence


Title: "Accelerating Data-Parallel Primitives and Multi-way Joins on
Heterogeneous Processors"

by

Mr. Zhuohang LAI


Abstract:

Data-parallel primitives, such as gather, scatter, scan (prefix sum), and 
split, are widely used in parallel programs. Multi-way joins are a common 
operator in data analytics applications. In this proposal, we design and 
implement efficient algorithms for these primitives and join operators on 
heterogeneous processors, including multi-core CPUs, Intel Xeon Phi (KNC) 
processors, and Graphics Processing Units (GPUs).

Specifically, we first revisit the performance of scatter and gather on 
new-generation GPUs, and propose a new model for their optimization. We then 
propose optimization strategies for these two primitives as well as scan and 
split that work well for an Intel multi-core CPU, an NVIDIA GPU, and a KNC. 
Finally, we propose a GPU-based multi-way hash join solution that effectively 
utilizes the primitives to achieve high bandwidth utilization on GPUs.


Date:			Friday, 19 June 2020

Time:                  	2:00pm - 4:00pm

Zoom Meeting:		https://hkust.zoom.us/j/97539050459

Committee Members:	Dr. Qiong Luo (Supervisor)
  			Dr. Wei Wang (Chairperson)
 			Prof. Ke Yi
 			Dr. Wei Zhang (ECE)


**** ALL are Welcome ****