Sort vs. Hash Join Revisited for Near-Memory Execution

Data movement between memory and CPU is a well-known energy bottleneck for analytics. Near-Memory Processing (NMP) is a promising approach for eliminating this bottleneck by shifting the bulk of the computation toward memory arrays in emerging stacked DRAM chips. Recent work in this space has been limited to regular computations that can be localized to a single DRAM partition. This paper examines a Join workload, which is fundamental to analytics and is characterized by irregular memory access patterns. We consider several join algorithms and show that while near-data execution can improve both energy-efficiency and performance, effective NMP algorithms must consider locality, access granularity, and microarchitecture of the stacked memory devices.

Publié dans:
5th Workshop on Architectures and Systems for Big Data (ASBD)
Présenté à:
5th Workshop on Architectures and Systems for Big Data (ASBD 2015), Portland, Oregon, USA, June 13, 2015

Note: Le statut de ce fichier est: Anyone

 Notice créée le 2015-06-22, modifiée le 2020-07-29

Publisher's version:
Télécharger le document

Évaluer ce document:

Rate this document:
(Pas encore évalué)