‘GPU-accelerated Multiphysics Simulation’

In recent technology developments General Purpose computation on Graphics Processor Units (GPGPU) has been recognized a viable HPC technique. In this context, GPUacceleration is rooted in high-order Single Instruction Multiple Data (SIMD)/Single Instruction Multiple Thread (SIMT) vector-processing capability, combined with highspeed asynchronous I/O and sophisticated parallel cache memory architecture. In this presentation we examine the enParallel, Inc. (ePX) approach in leveraging this technology for accelerated multiphysics computation {1}{2}{3}. As is well understood, both complexity and size impact realizable multiphysics simulation performance. Multiphysics applications by definition incorporate diverse model components, each of which employs characteristic algorithmic kernels, (e.g. sparse/dense linear solvers, gradient optimizers, multidimensional FFT/IFFT, wavelet, random variate generators). This complexity is further increased by any requirement for structured communications across module boundaries, (e.g. dynamic boundary conditions, multi-grid (re)discretization, and management of disparate time-scales). Further, multiphysics applications tend toward large scale and long runtimes due to; (a) presence of multiple physical processes and (b) high-order discretization as result of persistent nonlinearity, chaotic dynamics, etc. It then follows acceleration is highly motivated, and any associated performance optimization schema must be sufficiently sophisticated so as to address all salient aspects of process resource mapping and scheduling, and datapath movement. For the GPU-accelerated cluster, this remains a particularly important consideration due to the fact GPU lends an additional degree of freedom to any choice of processing resource; multiphysics performance optimization then reduces to a goal of achieving highest possible effective parallelism across all available HPC resources, each of which is associated with a characteristic process model. In ePX applications, processing models are organized hierarchically so as to structurally minimize high-overhead interprocess communications; process optimization is then performed based upon an assumed scatter-gather principle recursively applied at distributed (cluster) and Symmetric Multi-Processor (SMP; multicore CPU) hierarchy levels. This approach supports flexible optimization across all physics modules. In particular, explicit pipelining of cluster, CPU, and GPU processes is implemented based upon asynchronous transaction calls at an associated Application Programming Interface (API). This generally improves effective parallelization beyond what might otherwise be possible. Further, a complete multiphysics application must be accelerated consistent with dictates of Amdahl’s Law. In this context, ePX is shown to exhibit a full-featured supercomputer processing model, highly optimized for GPU-accelerated clusters or workstations, and particularly well suited to multiphysics applications. The ePX framework is then presented as a generic and reusable multiphysics development solution featuring scatter-gather infrastructure as a software architectural component and obviating any need for specialized compilation technology or OS runtime support.