An OpenSHMEM Implementation for the Adapteva Epiphany Coprocessor

James Ross1(B) and David Richie2

1 U.S. Army Research Laboratory,

Aberdeen Proving Ground, MD 21005, USA This email address is being protected from spam bots, you need Javascript enabled to view it 2 Brown Deer Technology, Forest Hill, MD 21050, USA This email address is being protected from spam bots, you need Javascript enabled to view it

Abstract. This paper reports the implementation and performance evaluation of the OpenSHMEM 1.3 specification for the Adapteva Epiphany architecture within the Parallella single-board computer. The Epiphany architecture exhibits massive many-core scalability with a physically compact 2D array of RISC CPU cores and a fast network-on- chip (NoC). While fully capable of MPMD execution, the physical topology and memory-mapped capabilities of the core and network translate well to Partitioned Global Address Space (PGAS) programming models and SPMD execution with SHMEM.

Keywords: OpenSHMEM • Network-on-chip (NoC) • Single-board computer • Performance evaluation

Introduction and Motivation

The OpenSHMEM communications library is designed for computer platforms using Partitioned Global Address Space (PGAS) programming models [1]. Historically, these were large Cray supercomputers, but now the OpenSHMEM interface may also be used on commodity clusters. The Adapteva Epiphany architecture represents a divergence in computer architectures typically used with OpenSHMEM and is just one of many emerging parallel architectures that present a challenge in identifying effective programming models to exploit them. While some researchers may be considering how the OpenSHMEM API may interact with coprocessors, the work presented here leverages the API for devicelevel operation. In some aspects, the Epiphany architecture resembles a symmetric multiprocessing (SMP) multi-core processor with a shared off-chip global memory pool. However, each core can directly address the private address space of neighboring cores across an on-chip 2D mesh network. Thus, the architecture also has the characteristics of a PGAS platform. Previous proof-of-concepts demonstrated that message passing protocols could achieve good application performance on the Epiphany architecture [2,3]. However, it was unclear if the © Springer International Publishing AG 2016

M. Gorentla Venkata et al. (Eds.): OpenSHMEM 2016, LNCS 10007, pp. 146-159, 2016. DOI: 10.1007/978-3-319-50995-2.10

OpenSHMEM 1.3 standard could be fully implemented within the platform limitations and achieve high performance using a standard programming model without resorting to non-standard software extensions.

Existing open source OpenSHMEM implementations are inadequate within the constraints of the Epiphany architecture, so a new C language implementation named ARL OpenSHMEM for Epiphany was developed from scratch. The design emphasizes a reduced memory footprint, high performance, and simplicity, which are often competing goals. This paper discusses the Epiphany architecture in Sect. 2.1, the OpenSHMEM implementation and performance evaluation in Sect. 3, and a discussion of future work and potential standard extensions for embedded architectures in Sect. 4.

< Prev   CONTENTS   Source   Next >