FPGA-BASED HYBRID COMPUTING FOR ESS LINAC SIMULATOR.

Abstract

The thesis explores efficient implementation strategies for the European Spallation Source (ESS) linear accelerator simulator. The target simulator needs to run at real time, requires high computation accuracy, and should be scalable for high density beam scenarios. The high data processing, communication, and storage requirements due to a large data set, along with a strict accuracy requirement, poses a critical implementation challenge for traditional computing platforms.

To tackle these issues, this thesis uses a scalable platform with hybrid computing capabilities and the OpenCL framework for an unified programming model. The hybrid computing platform allows for mapping tasks to the most suitable hardware and explores heterogeneous memory hierarchy to fast data shuffling. The OpenCL framework allows functional portability and scalability across different target devices such as CPU, GPU and OpenCL accelerator devices like with FPGA and DSP arrays. The computational intensive tasks of the simulator can be conveniently mapped to the accelerators, where computational parallelism is explored.

The targeted simulator is implemented in a Xilinx hybrid computing platform, consisting of an Intel i7 CPU, an Nvidia 960 GPU, and a Xilinx Kintex Ultra-scale FPGA. Comparing to the benchmark (a C++ based implementation), we are able to accelerate the ESS simulator by more than 80x on the GPU and 25x with FPGA, with the same simulation accuracy (double precision floating point). We identified the implementation bottleneck on the specific platform, which is the memory bandwidth. This leads to our future work. One important future task is to investigate different hybrid computing platforms of different vendors, considering computation capability, memory bandwidth, as well as design software. Moreover, different data types will be examined, including fixed-point, double/single-precision floating point, or custom floating point.

Popular Abstract

The European Spallation Source is one of Europe's largest research infrastructures to bring new insights into the challenges of science and innovation in fields as diverse as material and life sciences, energy, environmental technology, cultural heritage,solid-state and fundamental physics by the end of the decade. A 5 MW, long pulse proton accelerator is used to reach this goal. The particles in this beam contains so much of energy, that the loses to the super-conducting structures become marginally critical. An accurate and real-time simulation model to predict the behaviour of the particles is required to minimize the losses. There are two approaches to model the particle behaviour, one is to use an approximation with the envelop of the beam and the other is to model each particle and track them. The envelop method can run real-time, but has discrepancies to the actual behaviour of the beam at high energy configurations, hence a computationally intensive mult-particle simulation model is required to provide realistic behaviour. This is achieved by supporting non-linear behaviour such as space charge (particle to particle interactions). The multi-particle simulation algorithm has linear and non-linear attributes, where the linear section can map efficiently to the graphic processor hardware architecture, while the non-linear section can map more efficiently to the flexible FPGA fabrics. To make the multi-particle simulation model run real-time, a computing platform that can support graphic processors and FPGA accelerators are required. The thesis proposes to use a hybrid computing platform with FPGA, GPU and CPU to provide the efficient mapping of the multi-particle simulation model. A unified development environment and a scalable platform is developed by using a software stack build based on OpenCL framework. This allowed to develop the system in a rapid design phase and maximize the code re-use. We were able to accelerate the linear part of the simulation model by upto 89x on the given set of hardware. Different ways to further optimize the solution are also explored.

• author: Jeevaraj, Arun
• supervisor: Liang Liu; Fredrik Edman; Emanuele Laface; Maurizio Donna
• organization: Department of Electrical and Information Technology
• alternative title: FPGA-baserad hybridberäkningsenhet för simulering av ESS linjäraccelerator.
• course: EITM02 20171
• year: 2017
• type: H2 - Master's Degree (Two Years)
• subject: Technology and Engineering
• keywords: Hybrid Computing, OpenCL, FPGA, Linear Accelerator, Heterogeneous computing
• report number: LU/LTH-EIT 2017-609
• language: English
• id: 8928453
• date added to LUP: 2017-11-13 11:57:00
• date last changed: 2017-11-13 11:57:00