This paper presents a scalable and high-Throughput hybrid computing platform for the real-Time multi-particle based Linac (Linear accelerator) simulation model to be used at the European Spallation Source (ESS). The multi-particle simulation model with non-linear modeling is needed to provide a realistic behavior of the particle beam for reducing the losses at the superconducting structures. The computation complexity of the simulations can reach 1012 matrix multiplication operations for a test case of 106 beam particles simulated over 106 cells. An OpenCL (Open Computing Language) based framework is used to map the processing intensive parts of the simulation model efficiently to any configuration of a CPU-, GPU-and FPGA-based platform. Optimizations using data precision strategies have also been explored to further improve the throughput after reaching memory access saturation. We are able to achieve up to 89 × speed up compared to a C++ benchmark of the same system.
|Title of host publication
|2018 IEEE Nordic Circuits and Systems Conference, NORCAS 2018
|Subtitle of host publication
|NORCHIP and International Symposium of System-on-Chip, SoC 2018 - Proceedings
|Jari Nurmi, Peeter Ellervee, Juri Mihhailov, Kalle Tammemae, Maksim Jenihhin
|IEEE - Institute of Electrical and Electronics Engineers Inc.
|Published - 2018 Dec 11
|4th IEEE Nordic Circuits and Systems Conference, NORCAS 2018: NORCHIP and International Symposium of System-on-Chip, SoC 2018 - Tallinn, Estonia
Duration: 2018 Oct 30 → 2018 Oct 31
|4th IEEE Nordic Circuits and Systems Conference, NORCAS 2018: NORCHIP and International Symposium of System-on-Chip, SoC 2018
|2018/10/30 → 2018/10/31
Subject classification (UKÄ)
- Computer Science