Marc Snir

Minyu Cheng, Jiakun Yan, Marc Snir

The bulk synchronous parallel (BSP) model struggles with irregular workloads due to rigid global communication. While fine-grained asynchronous BSP (FA-BSP) improves overlap, existing implementations typically rely on a limiting one-process-per-core model. This paper proposes a multithreaded FA-BSP approach combining Lightweight Communication Interface (LCI) and OpenMP to fully exploit multicore architectures. We evaluate this design using the NAS Parallel Benchmark Integer Sort (IS), retaining the original irregular Gaussian distribution to rigorously test load balancing. By replacing synchronous MPI collectives with OpenMP multithreading and LCI's fine-grained, zero-copy active messages, we enable efficient computation-communication overlap. Our evaluation demonstrates that multithreaded FA-BSP significantly outperforms traditional bulk-synchronous MPI implementations, offering a scalable solution for irregular scientific applications.

Nikoli Dryden, Naoya Maruyama, Tom Benson et al.

Scaling CNN training is necessary to keep up with growing datasets and reduce training time. We also see an emerging need to handle datasets with very large samples, where memory requirements for training are large. Existing training frameworks use a data-parallel approach that partitions samples within a mini-batch, but limits to scaling the mini-batch size and memory consumption makes this untenable for large samples. We describe and implement new approaches to convolution, which parallelize using spatial decomposition or a combination of sample and spatial decomposition. This introduces many performance knobs for a network, so we develop a performance model for CNNs and present a method for using it to automatically determine efficient parallelization strategies. We evaluate our algorithms with microbenchmarks and image classification with ResNet-50. Our algorithms allow us to prototype a model for a mesh-tangling dataset, where sample sizes are very large. We show that our parallelization achieves excellent strong and weak scaling and enables training for previously unreachable datasets.