 |
||
|
||
Home > TERATEC FORUM > Workshop 6
Algorithms and parallel software
Scalability & performances of general purpose fluid dynamics solver on low power cluster: new perspectives on combined CUDA-ARM architecture
Gino PERNA, ENGINSOFT SpA
With the arrival of 64-bit ARM as well as new host 'REAL" platforms and HPC H2020 projects ready to start, we had the opportunity to study and gets hands-on on one of the first of this architectures in a real cluster connected via FDR infiniband, coupled with GPU.
In order to test the real capabilities in engineering field and as well because involved in many CFD and HPC projects we have perfomed a complete porting of a CFD code with a computational engine based on GPU in order to prove scalability even on low power CPU's. The CFD code is Sailfish, a free computational fluid dynamics solver based on the Lattice Boltzmann method and optimized for modern multi-core systems, especially GPUs (Graphics Processing Units). The hardware solution used are based on ARMv8 SoC plus Nvidia GPUs and interconnected with infiniband FDR fabric. In detail, the first machine has single SoC X-gene1 by APM, with 8 cores ARMv8 2.4 GHz with 16 PCIe Gen3 lines 8 connected to the GPU, and 8lines to the infiniband card. A four node cluster has been used as a test environment. The second system is based on CAVIUM Thunder-X Soc, 48 cores ARMv8, single node.
This article presents the work done and performance data obtained.
 |
Gino Perna is HPC and It Manager at Enginsoft. He obtained his MSc degree in Civil Engineering at Padua University in 1986 with a Master thesis on computer aided design of Dams. In Enginsoft since 1987 he occupied different positions in CAE Simulation/Consulting and Verticalization of procedures for the CAE market. Last ten years have been dedicated to HPC etherogeneous computational and visualization clusters maintained by his staff and to help and assist big companies in running and maintaining private clusters.
In addition to his duties at Enginsoft he has teaching duties at the University of Trento, mostly focused on Computer Programming. His knowledge goes through Mechanical & CFD simulations with last few years spent in the field of HPC and CAE. |
