MSC Nastran delivers scalable performance when using InfiniBand and therefore an efficient tool for manufacturing vendors MSC Nastran MPI Profiling Profiling the application is essential for understanding its performance dependency on the various cluster subsystems. In particular, application communication profiling can help in choosing the most efficient interconnect and MPI library, and in identifying the critical communication sensitivity points that greatly influence the application’s performance, scalability and productivity.
MPI Library Comparisons We have compared the performance of two MPI libraries – Open MPI and HP MPI. The results are presented in figure 4. HP MPI shows slightly higher performance on larger number of nodes compared to Open MPI, while Open MPI performs better on smaller number of nodes. In order to achieve the best performance with Open MPI, we have modified Open MPI to allow InfiniBand support - the openib BTL was not built with the Open MPI that is being shipped with MSC Nastran. Processor binding was also used to enhance performance with the Open MPI MCA using “mpi_paffinity_alone 1”.
Conclusions From concept to engineering and from design to test and manufacturing, engineering relies on powerful virtual development solutions. Finite Element Analysis (FEA) and Computational Fluid Dynamics (CFD) are used in an effort to secure quality and speed up the development process. Cluster solutions maximize the total value of ownership for FEA and CFD environments and extend innovation in virtual product development.
MSC Nastran MPI profiling data (Figures 2 and 3), show the breakdown of MPI calls, the MPI message sizes being used and the time spent by MPI calls in several cluster HPC cluster environments impose configurations. high demands for cluster connectivity MPI_Ssend and MPI_Recv are the most throughput, low-latency, low CPU overhead, used MPI calls with the MSC Nastran network flexibility and high-efficiency in xl0tdf1 case (MPI_Ssend is a blocking order to maintain a balanced system and synchronized send). Each of these MPI functions accounts for nearly half of all MPI functions. The xl0tdf1 shows some data communication as well, but not as large as the synchronization calls. From Figure 3 we learn that the majority of MPI messages are small messages, and mostly fall in the range between 0 and 64 bytes. Small message sizes are typically used Figure 1: Interconnect comparison with xl0tdf1 for synchronization and large message sizes are typically used for data communication (Send/Recv). The MSC Nastran xl0tdf1 case spends the majority of the time on computation and less on communications. Still, the interconnect influence is noticeable and can greatly influence the application and the Figure 3: MSC Nastran Profiling – cluster productivity. Number of MPI Calls
to achieve high application performance and scaling. Low-performance interconnect solutions, or lack of interconnect hardware capabilities will result in degraded system and application performance. MSC Nastran software was investigated in this paper. In all InfiniBand-based cases, MSC Nastran demonstrated high parallelism and scalability, which enabled it to take full advantage of multi-core HPC clusters. Moreover, according to the results, a lower-speed interconnect, such as Ethernet becomes ineffective on large cluster sizes, and can cause a reduction in performance. We have profiled the communication over the network of MSC Nastran software to determine its sensitivity points, which is essential in order to estimate the influence of the various cluster components, both hardware and software. We have provided the key points on how to configure the MPI in order to achieve optimum performance and scalability. We would like to thank the HPC Advisory Council for providing the resources for completing this research and testing. We would like to thank MSC for the tight collaborations and for providing the licenses for the software. The testing results demonstrated the MSC Nastran delivers scalable performance when using InfiniBand and therefore an efficient tool for manufacturing vendors. u
Figure 2: MSC Nastran Profiling – Number of MPI Calls
Figure 4: Time Spent by MPI Volume II - Winter 2012
| 31