Near-Memory Data Services

EDIC candidacy exam
Exam president: Prof. James Larus
Thesis advisor: Prof. Babak Falsafi
Co-examiner: Prof. Paolo Ienne

Abstract
Modern data analytics operate on huge in-memory datasets. The high-energy cost of moving data from memory to CPU has become a major concern with the end of Dennard scaling. Near-memory processing (NMP) has become an interesting direction for data analytics to reduce the cost of data movement, especially after breakthrough in 3D-Stack technology that allows tight integration of logic and memory. We present three different design alternatives for NMP data analytics. The designs focus on the tradeoffs between programmability and performance under tight NMP constraints. We discuss each design's architecture, advantages, and limitations. We finally propose a plan that builds upon the previous work to produce a high performance NMP system that targets various data analytics workloads.

Background papers
Practical Near Data Processing ,Mingyu Gao, Grant Ayers, Christos Kozyrakis, Stanford edu.
HRL: Efficient and Flexible Reconfigurable Logic for Near-Data Processing, Mingyu Gao and Christos Kozyrakis, Stanford edu
The Mondrian Data Engine, Drumond E., et al.