Transformer models trained on long sequences often achieve higher accuracy than short sequences. Unfortunately, conventional transformers struggle with long sequence training due to the overwhelming computation and memory requirements. Existing methods for long sequence training offer limited speedup and memory reduction, and may compromise accuracy. This paper presents a novel and efficient distributed training method, the Long Short-Sequence Transformer (LSS Transformer), for training transformer with long sequences. It distributes a long sequence into segments among GPUs, with each GPU computing a partial self-attention for its segment. Then, it uses a fused communication and a novel double gradient averaging technique to avoid the need to aggregate partial self-attention and minimize communication overhead. We evaluated the performance between LSS Transformer and the state-of-the-art Nvidia sequence parallelism on a Wikipedia enwik8 dataset. Results show that our proposed method lead to 5.6x faster and 10.2x more memory-efficient implementation compared to state-of-the-art sequence parallelism on 144 Nvidia V100 GPUs. Moreover, our algorithm scales to an extreme sequence length of 50,112 at 3,456 GPUs, achieving 161% super-linear parallel efficiency and a throughput of 32 petaflops.
Ptychography is a popular microscopic imaging modality for many scientific discoveries and sets the record for highest image resolution. Unfortunately, the high image resolution for ptychographic reconstruction requires significant amount of memory and computations, forcing many applications to compromise their image resolution in exchange for a smaller memory footprint and a shorter reconstruction time. In this paper, we propose a novel image gradient decomposition method that significantly reduces the memory footprint for ptychographic reconstruction by tessellating image gradients and diffraction measurements into tiles. In addition, we propose a parallel image gradient decomposition method that enables asynchronous point-to-point communications and parallel pipelining with minimal overhead on a large number of GPUs. Our experiments on a Titanate material dataset (PbTiO3) with 16632 probe locations show that our Gradient Decomposition algorithm reduces memory footprint by 51 times. In addition, it achieves time-to-solution within 2.2 minutes by scaling to 4158 GPUs with a super-linear speedup at 364% efficiency. This performance is 2.7 times more memory efficient, 9 times more scalable and 86 times faster than the state-of-the-art algorithm.
Scientific communities are increasingly adopting machine learning and deep learning models in their applications to accelerate scientific insights. High performance computing systems are pushing the frontiers of performance with a rich diversity of hardware resources and massive scale-out capabilities. There is a critical need to understand fair and effective benchmarking of machine learning applications that are representative of real-world scientific use cases. MLPerf is a community-driven standard to benchmark machine learning workloads, focusing on end-to-end performance metrics. In this paper, we introduce MLPerf HPC, a benchmark suite of large-scale scientific machine learning training applications driven by the MLCommons Association. We present the results from the first submission round, including a diverse set of some of the world's largest HPC systems. We develop a systematic framework for their joint analysis and compare them in terms of data staging, algorithmic convergence, and compute performance. As a result, we gain a quantitative understanding of optimizations on different subsystems such as staging and on-node loading of data, compute-unit utilization, and communication scheduling, enabling overall $>10 \times$ (end-to-end) performance improvements through system scaling. Notably, our analysis shows a scale-dependent interplay between the dataset size, a system's memory hierarchy, and training convergence that underlines the importance of near-compute storage. To overcome the data-parallel scalability challenge at large batch sizes, we discuss specific learning techniques and hybrid data-and-model parallelism that are effective on large systems. We conclude by characterizing each benchmark with respect to low-level memory, I/O, and network behavior to parameterize extended roofline performance models in future rounds.
To address the unprecedented scale of HL-LHC data, the Exa.TrkX project is investigating a variety of machine learning approaches to particle track reconstruction. The most promising of these solutions, graph neural networks (GNN), process the event as a graph that connects track measurements (detector hits corresponding to nodes) with candidate line segments between the hits (corresponding to edges). Detector information can be associated with nodes and edges, enabling a GNN to propagate the embedded parameters around the graph and predict node-, edge- and graph-level observables. Previously, message-passing GNNs have shown success in predicting doublet likelihood, and we here report updates on the state-of-the-art architectures for this task. In addition, the Exa.TrkX project has investigated innovations in both graph construction, and embedded representations, in an effort to achieve fully learned end-to-end track finding. Hence, we present a suite of extensions to the original model, with encouraging results for hitgraph classification. In addition, we explore increased performance by constructing graphs from learned representations which contain non-linear metric structure, allowing for efficient clustering and neighborhood queries of data points. We demonstrate how this framework fits in with both traditional clustering pipelines, and GNN approaches. The embedded graphs feed into high-accuracy doublet and triplet classifiers, or can be used as an end-to-end track classifier by clustering in an embedded space. A set of post-processing methods improve performance with knowledge of the detector physics. Finally, we present numerical results on the TrackML particle tracking challenge dataset, where our framework shows favorable results in both seeding and track finding.
* Proceedings submission in Connecting the Dots Workshop 2020, 10 pages