Speeding Up Vector Engine Offloading with AVEO

Erich Focht ()
Additional contact information
Erich Focht: NEC Deutschland GmbH

A chapter in Sustained Simulation Performance 2019 and 2020, 2021, pp 35-47 from Springer

Abstract: Abstract Vector Engine Offloading (VEO) was the first implementation of an API for programming the SX-Aurora Tsubasa Vector Engine (VE) like an accelerator, i.e. writing programs for the host CPU which call certain offloading kernels running on the VE. The native VE programming model using OpenMP and MPI still dominates in applications, but CUDA, HIP, OpenMP Target, OpenACC, OpenCL find more and more traction. This report introduces AVEO, an alternative VE offloading implementation with VEO compatible API. It was redesigned to solve a set of problems in VEO and improve call latency as well as memory transfer bandwidth. The results show latency improvements of up to factor 18 and bandwidth increases by factor 8–10 for small buffers and 15–20% for very large buffers. We describe implementation details and remote memory access mechanisms as well as API extensions. This development should contribute to making accelerator-style hybrid programming more attractive on the vector engine, ease porting of hybrid programs but also developing more sophisticated hybrid programming frameworks.

Date: 2021
References: Add references at CitEc
Citations:

There are no downloads for this item, see the EconPapers FAQ for hints about obtaining it.

Related works:
This item may be available elsewhere in EconPapers: Search for items with the same title.

Export reference: BibTeX RIS (EndNote, ProCite, RefMan) HTML/Text

Persistent link: https://EconPapers.repec.org/RePEc:spr:sprchp:978-3-030-68049-7_3

Ordering information: This item can be ordered from
http://www.springer.com/9783030680497

DOI: 10.1007/978-3-030-68049-7_3

Access Statistics for this chapter

More chapters in Springer Books from Springer
Bibliographic data for series maintained by Sonal Shukla () and Springer Nature Abstracting and Indexing ().