Understanding Data Analytics Workloads on Intel(R) Xeon Phi(R)
Paper i proceeding, 2016

The Intel® Xeon Phi™ is gaining popularity for high-performance computing (HPC) applications, but the performance of this many-core coprocessor with wide floating point SIMD units has yet to be explored on data analytics workloads. We construct a benchmark suite to explore the Xeon Phi™'s potential for use in data center servers. Our resulting PhiBench consists of eight representative data analytics workloads covering six application domains. These workloads are optimized for Xeon Phi™ and implemented with openMP and Cilk Plus. We run them on real-world datasets and compare their performances for different programming models, input data sizes, and thread counts. Most benefit from the Xeon Phi™'s high computational capacity, delivering speedups by factors of four to almost 29. We further analyze their microarchitecture-level performance characteristics, including vectorization intensities and cache behaviors, and we investigate the impact of affinities and scheduling policies on performance and scalability. Our observations should help other researchers and practitioners to understand and optimize the behaviors of data analytics workloads on the Xeon Phi™.

Performance characterization

Data analytics

Xeon Phi™

Författare

B. Xie

Chinese Academy of Sciences

X. Liu

College of William and Mary

Sally A McKee

Chalmers, Data- och informationsteknik, Datorteknik

J. Zhan

Chinese Academy of Sciences

Z. Jia

Chinese Academy of Sciences

L. Wang

Chinese Academy of Sciences

L. Zhang

Chinese Academy of Sciences

18th IEEE International Conference on High Performance Computing and Communications, 14th IEEE International Conference on Smart City and 2nd IEEE International Conference on Data Science and Systems, HPCC/SmartCity/DSS 2016, Sydney, Australia, 12-14 December 2016

206-215

Ämneskategorier

Datavetenskap (datalogi)

DOI

10.1109/HPCC-SmartCity-DSS.2016.0039

ISBN

978-1-5090-4296-8

Mer information

Skapat

2017-10-07