General Purpose High Performance Computing Research Articles

Investigating the Dirac Operator Evaluation with FPGAs

In recent years the computational capacity of single Field Programmable Gate Arrays (FPGA) devices as well as their versatility has increased significantly. Adding to that the High Level Synthesis frameworks allowing to program such processors in a high level language like C++, makes modern FPGA devices a serious candidate as building blocks of a general purpose High Performance Computing solution. In this contribution we describe benchmarks which we performed using a Lattice QCD code, a highly compute-demanding HPC academic code for elementary particle simulations. We benchmark the performance of a single FPGA device running in two modes: using the external or embedded memory. We discuss both approaches in detail using the Xilinx U250 device and provide estimates for the necessary memory throughput and the minimal amount of resources needed to deliver optimal performance depending on the available hardware platform.

Use of GPUs in Trigger Systems

Abstract In recent years the interest for using graphics processor (GPU) in general purpose high performance computing is constantly rising. In this paper we discuss the possible use of GPUs to construct a fast and effective real time trigger system, both in software and hardware levels. In particular, we study the integration of such a system in the NA62 trigger. The first application of GPUs for rings pattern recognition in the RICH will be presented. The results obtained show that there are not showstoppers in trigger systems with relatively low latency. Thanks to the use of off-the-shelf technology, in continous development for purposes related to video game and image processing market, the architecture described would be easily exported to other experiments, to build a versatile and fully customizable online selection.

OpenMP Optimization and its Translation to OpenGL

For general purpose high-performance computing, recently GPGPUs have emerged as powerful vehicles. Programming GPGPUs is complex when compared to programming general purpose CPUs and parallel programming models such as OpenMP. Goal of our translation is to improve programmability and make existing OpenMP applications to be able to execute on GPGPUs. OpenMP has established itself as an important method and language extension for programming shared-memory parallel computers. Our translator works well on regular applications, leading to performance improvements of up to 50X over the unoptimized translation.

From Silicon to Science

The field of high performance computing (HPC) currently abounds with excitement about the potential of a broad class of things called accelerators . And, yet, few accelerator based systems are being deployed in general purpose HPC environments. Why is that? This article explores the challenges that accelerators face in the HPC world, with a specific focus on FPGA based systems. We begin with an overview of the characteristics and challenges of typical HPC systems and applications and discuss why FPGAs have the potential to have a significant impact. The bulk of the article is focused on twelve specific areas where FPGA researchers can make contributions to hasten the adoption of FPGAs in HPC environments.

OpenMP to GPGPU

GPGPUs have recently emerged as powerful vehicles for general-purpose high-performance computing. Although a new Compute Unified Device Architecture (CUDA) programming model from NVIDIA offers improved programmability for general computing, programming GPGPUs is still complex and error-prone. This paper presents a compiler framework for automatic source-to-source translation of standard OpenMP applications into CUDA-based GPGPU applications. The goal of this translation is to further improve programmability and make existing OpenMP applications amenable to execution on GPGPUs. In this paper, we have identified several key transformation techniques, which enable efficient GPU global memory access, to achieve high performance. Experimental results from two important kernels (JACOBI and SPMUL) and two NAS OpenMP Parallel Benchmarks (EP and CG) show that the described translator and compile-time optimizations work well on both regular and irregular applications, leading to performance improvements of up to 50X over the unoptimized translation (up to 328X over serial).