Altair Optistruct* Intel® Xeon® Processor E5-2600 v4 Product Family Intel® SSD Data Center Family for PCIe* white paper

Speeding up Altair OptiStruct* Simulations with the Intel® SSD Data Center Family for PCIe* Altair OptiStruct* provides engineers and designers with a unified solution from concept to final design by leveraging advanced analysis capabilities and novel, optimization-driven simulation. In this process, the simulation time for one optimization iteration is a critical consideration, because it affects the computational speed and scalability of the entire design process. Authors Hongwei Zhou Altair Engineering, Inc.: hongwei@altair.com Nick Meng Intel Corporation: nick.meng@intel.com With ongoing advances in processing power for high-performance computing, a bottleneck has emerged in the memory domain. Consequently, the typically low memory requirements of iterative solvers have made them attractive in many engineering simulation applications. However, for structural solid mechanics applications, the direct solver is still preferred over the iterative solver due to greater robustness and accuracy in handling large, complex industrial models. Large factor matrices are typically involved, which can be as large as a few hundred gigabytes; as a result, memory is the critical impediment for large engineering model simulations using direct solvers. While external memory is frequently used to circumvent this issue for large models, traditional SATA hard drives are limited by slow read/write rates, reducing the utility of out-of-core methods. Modern operating systems are typically capable of utilizing available system memory for I/O buffering, but while this practice is effective and commonly used, the computational price is high. Domain-decomposition techniques based on distributed memory programming (DMP) models using message passing interface (MPI) techniques can significantly reduce memory usage on individual computational nodes. This method offers the additional advantage of increased computational scalability. DMP is highly effective because it distributes CPU computation, memory access, and I/O activity. However, a balance must be maintained between placing more MPI ranks and retaining memory for I/O buffering. The emergence of solid-state drives (SSDs) raises sustained read/write performance, offering significant value for large-dataset I/O operations. SSDs are therefore well suited for OptiStruct simulations. This paper benchmarks two cases using Altair Optistruct 2017.0: an engine block nonlinear static simulation and a buckling analysis on a full-car model. The tests were conducted using a two-socket system based on Intel® Xeon® processors E5-2667 v4 @ 3.2 GHz and 256 GB PC2400 DDR4 RAM, with five different storage configurations: • Single hard-disk drive (HDD): 1x Seagate Constellation.2* ST91000640NS (1 TB) • Single SATA SSD: 1x Intel® SSD Data Center (DC) S3710 Series (800 GB) • Single NVMe SSD: 1x Intel SSD DC P3600 Series (2 TB) • RAID0, SATA SSDs: 4x Intel SSD DC S3710 Series (800 GB) • RAID0, NVMe SSDs: 3x Intel SSD DC P3600 Series (1.6 TB) White Paper | Speeding up Altair OptiStruct* Simulations with the Intel® SSD Data Center Family for PCIe* Case 1: Engine Block Nonlinear Static Analysis Case 1a: Overall Time on Nonlinear Analysis Engine models typically consist of solid tetrahedral or rs)30.0 hexahedral elements, generally leading to higher bandwidth Hou 25.0 in the sparse matrix and introducing a tremendously large ( e20.0 factor matrix. For the 20 million degrees of freedom model, Tim 15.0 storing the factor matrix and internal stack, in addition to ed other working storage, requires more than 300 GB of space. 10.0 sp In such cases, the sparse solver dumps the whole factor Ela 5.0 matrix onto the disk and retrieves it during the backward and 0.0 forward substitution