Cost-effective Sizing of your HPC Cluster for CAE Simulations

The use of CAE techniques as an integral part of product development has become indispensable in the automotive industry. The investigation of different problems with the help of virtual simulation is used daily by development teams. They need considerable computing power, which is provided by HPC (High Performance Computing) clusters. Typical cluster installations in the automotive industry can range from local resources with 50 to 200 cores up to centralized cloud computing installations with more than 10,000 cores. However, in general, the requirements for the HPC systems are higher than available power. This has economic (limitation of licenses) and infrastructural (limited cooling facilities and space in the data center) reasons. Therefore, an optimized schedule of the calculations in batch operation is essential in order to achieve maximum efficiency of calculations. GNS Systems is a service provider specializing in these kinds of challenges. In this contribution we would like to give examples about how we achieve more efficiency by detailed and scientific analysis of the situation given by the customer. First, using benchmark results on modern Intel Xeon E5 processors, we investigate how the performance of various CAE software depends on the number of CPU cores used and on the details of the CPU core binding. The positive effect of the turbo frequency can be clearly observed. Second, we derive a general formula for the cost of a simulation (including hardware, software, and personnel cost). We show that in situations that are typical for normal commercial users in the automotive industry, there is always a "sweet spot" of CPU cores, i.e., a number of CPU cores other than 1 at which the total cost of a simulation is minimized. Last, we discuss possible tricks (such as underloading of hardware) which a typical industry user could make use of in order to reduce the total cost of simulations while leaving the total capacity of his cluster unchanged.