Dramatic growth in the use of computing in educational practices and research workflows encourages two-year institutions to adopt large-scale computing practices rapidly. Advanced cyberinfrastructure (CI) resources are required to support educators, students, and researchers in cloud computing, data sciences, and related cross-cutting fields such as smart manufacturing. While community colleges will differ in enrollment, geographical location, business models, and programs offered, the underlying institutional needs to store, access, manage, analyze, compute, and curate their data remain the same. Working with regional efforts to advance computing and networking, we discuss collaborative models to develop appropriate adoption models with them.
Modern powerful accelerators and composable infrastructures put our simulation frameworks to the test. We will show that the acceleration of a simulation framework is absolutely critical for good performance and scaling. Building on our previous work using research software as a benchmark for computing clusters, the High Performance Research Computing Group (HPRC)1 compares the Kokkos and GPU acceleration packages of LAMMPS (Large-scale Atomic/Molecular Massively Parallel Simulator) molecular dynamics software with NVIDIA H100 and Intel Data Center GPU Max 1100 accelerators on the composable ACES cyber-infrastructure at Texas A&M University. We observe different computational and communication patterns emerge from the different codes, which in turn result in different performance and scaling characteristics across these accelerators. We observe an opportunity for growth in the synergy between Intel’s oneAPI toolchain and the Kokkos framework to enable effective scaling of molecular dynamics simulations on composable infrastructure.
The potential for infinite scaling, improved performance, and better sharing of computing resources motivates researchers to adapt to composable infrastructures. Measuring performance on these systems requires an assessment of how the composed configuration itself affects performance which goes beyond traditional scaling approaches. We emphasize the subtle relationship between the nature of the calculation and the configuration of the composable infrastructure. New application benchmarking strategies are explored to inform the optimal configurations and best computing practices for composable systems. Realistic molecular dynamics research workflows are employed as benchmarks for composed GPU systems to develop a benchmarking strategy that yields recognizable and informative results. We employ the practices on a realistic case for a molecular dynamics research workflow on a composed GPU system. We discuss the identification of computational bottlenecks and establish a need for new benchmark performance suites that can help researchers articulate optimum compositions for composable infrastructure.