Development of an equation-based parallelization method for multiphase particle-in-cell simulations

Manufacturers have been developing new graphics processing unit (GPU) nodes with large capacity, high bandwidth memory and very high bandwidth intra-node interconnects. This enables moving large amounts of data between GPUs on the same node at low cost. However, small packet bandwidths and latencies have not decreased, which makes global dot products expensive. These characteristics favor a new kind of problem decomposition called equation decomposition rather than traditional domain decomposition. In this approach, each GPU is assigned one equation set to solve in parallel so that the frequent and expensive dot product synchronization points in traditional distributed linear solvers are eliminated. In exchange, the method involves infrequent movement of state variables over the high bandwidth, intra-node interconnects. To test this theory, our flagship code Multiphase Flow with Interphase eXchanges (MFiX) was ported to TensorFlow. This new product is known as MFiX-AI and can produce near identical results to the original version of MFiX with significant acceleration in multiphase particle-in-cell (MP-PIC) simulations. The performance of a single node with 4 NVIDIA A100s connected over NVLINK 2.0 was shown to be competitive to 1000 CPU cores (25 nodes) on the JOULE 2.0 supercomputer, leading to an energy savings of up to 90%. This is a substantial performance benefit for small- to intermediate-sized problems. This benefit is expected to grow as GPU nodes become more powerful. Further, MFiX-AI is poised to accept native artificial intelligence/machine learning models for further acceleration and development.

Automated summary

Manufacturers are developing new GPU nodes that have large capacity, high bandwidth memory, and very high bandwidth intra-node interconnects. This allows for low-cost data movement between GPUs on the same node. However, the expensive global dot products due to small packet bandwidths and latencies can be mitigated by using equation decomposition instead of traditional domain decomposition. Testing this theory, the code MFiX was ported to TensorFlow and resulted in the accelerated MFiX-AI, which showed competitive performance to a supercomputer with 1000 CPU cores, achieving significant energy savings.