Zero Redundancy Optimizer (ZeRO) has been used to train a wide range of large language models on massive GPU clusters due to its ease of use, efficiency, and good scalability. However, when training on low-bandwidth clusters, and/or when small batch size per GPU is used, ZeRO’s effective throughput is limited due to communication overheads. To alleviate this limitation, this paper introduces ZeRO++ composing of three communication volume reduction techniques (lowprecision all-gather, data remapping, and low-precision gradient averaging) to significantly reduce the communication volume up to 4x that enables up to 2.16x better throughput at 384 GPU scale. Our results also show ZeRO++ can speedup the RLHF by 3.3x compared to vanilla ZeRO. To verify the convergence of ZeRO++, we test up to 13B model for pretraining with 8/6-bits all gather and up to 30B model for finetuning with 4/2-bits all gather, and demonstrate on-par accuracy as original ZeRO (aka standard training). As a byproduct, the model trained with ZeRO++ is naturally weight-quantized, which can be directly used for inference without post-training quantization or quantization-aware training.

Computation in a typical Transformer-based large language model (LLM) can be characterized by batch size, hidden dimension, number of layers, and sequence length. Until now, system works for accelerating LLM training have focused on the first three dimensions: data parallelism for batch size, tensor parallelism for hidden size, and pipeline parallelism for model depth or layers. These widely studied forms of parallelism are not targeted or optimized for long sequence Transformer models. Given practical application needs for long sequence LLM, renewed attentions are being drawn to sequence parallelism. However, existing works in sequence parallelism are constrained by memory-communication inefficiency, limiting their scalability to long sequence large models. In this work, we introduce Ulysses, a novel, portable, and effective methodology for enabling highly efficient and scalable LLM training with extremely long sequence length. Ulysses at its core partitions input data along the sequence dimension and employs an efficient all-to-all collective communication for attention computation. Theoretical communication analysis shows that, whereas other methods incur communication overhead as sequence length increases, Ulysses maintains constant communication volume when sequence length and compute devices are increased proportionally. Furthermore, experimental evaluations show that Ulysses scales to more than 1 million context length and trains 2.5x faster with 4x longer sequence length than the existing method SOTA baseline.

Graph neural networks (GNNs) are a powerful approach for machine learning on graph datasets. Such datasets often consist of millions of modestly-sized graphs, making them well-suited for data-parallel training. However, existing methods show poor scaling due to load imbalances and kernel overheads. We propose an optimized 2D scatter-gather based represen-tation of GNNs that is amenable to distributed, data-parallel training without changing the underlying mathematics of the GNN. By padding graph data to a fixed size on each process, we can simplify data ingestion, make use of efficient compute kernels, equally distribute computation load, and reduce overheads. We benchmark edge-conditioned GNNs with the PCQM4M-LSC and OGB-PPA datasets. Our implementation shows better runtime performance than the state-of-the-art, with a$12\times$strong-scaling speedup on 16 GPUs and an$89.4\times\ \text{weak}$-scaling speedup on 100 GPUs.

Composite science workflows are gaining traction to manage the combined effects of (1) extreme hardware heterogeneity in new High Performance Computing (HPC) systems and (2) growing software complexity – effects necessitated by the convergence of traditional HPC with data sciences. Composing, analyzing, and optimizing a composite workflow remains highly challenging as the component technologies are generally developed in isolation and often feature widely varying levels of performance, scalability, and interoperability. In this paper, we propose novel workflow composition and analysis techniques to create and optimize a scalable and effective composite workflow for heterogeneous HPC centers, and define the performance space of variables that impact composite workflow performance. We present PerfFlowAspect, an Aspect Oriented Programming (AOP)-based tool to perform cross-cutting performance analysis of composite workflows and better understand the impact of key performance variables on workflows. Our solution directly addresses AOP concerns that can affect workflow performance and covers the full software lifecycle, ranging from the workflow's initial composition through performance analysis and optimization. We use our science workflow composition techniques to implement the American Heart Association Molecule Screening (AHA MoleS) workflow. Through experimentation, we demonstrate that tuning a single performance variable can improve AHA MoleS workflow performance by a factor of up to 2.45x. Our evaluation suggests that our techniques can significantly enhance the ability of a multi-disciplinary research and development team to create a high performance composite workflow.

Transformer models have revolutionized the field of Natural Language Processing (NLP) and they achieve state-of-the-art performance in applications like machine translation, question answering, regression, and summarization. However, training Transformers is challenging because of their large memory and compute requirements. The literature contains several approaches to parallelize training, like layer parallelism and pipeline parallelism, but they are optimized to benefit out-of-core models and they don't exploit the inherent parallelism in Transformer models. Other work uses model parallelism to achieve weak scaling by increasing the model size. In this paper, we propose sub-graph parallelism that provides a significant performance improvement over pure data parallelism with a fixed number of resources, and as an additional technique for strong- and weak-scaling without increasing model capacity. Our technique accelerates the training of Transformer models and we generalize the concept to any neural network with multiple branches. We optimize the communication for sub-graph parallelism and combine it with data parallelism to scale performance up to 1024 GPUs. To decrease communication overheads, we propose a topology-aware scheme that limits inter-node communication. Finally, we empirically compare sub-graph parallelism with pure data parallelism and demonstrate its performance benefits in end-to-end training.

With the rapid adoption of machine learning techniques for large-scale applications in science and engineering comes the convergence of two grand challenges in visualization. First, the utilization of black box models (e.g., deep neural networks) calls for advanced techniques in exploring and interpreting model behaviors. Second, the rapid growth in computing has produced enormous datasets that require techniques that can handle millions or more samples. Although some solutions to these interpretability challenges have been proposed, they typically do not scale beyond thousands of samples, nor do they provide the high-level intuition scientists are looking for. Here, we present the first scalable solution to explore and analyze high-dimensional functions often encountered in the scientific data analysis pipeline. By combining a new streaming neighborhood graph construction, the corresponding topology computation, and a novel data aggregation scheme, namely topology aware datacubes, we enable interactive exploration of both the topological and the geometric aspect of high-dimensional data. Following two use cases from high-energy-density (HED) physics and computational biology, we demonstrate how these capabilities have led to crucial new insights in both applications.

Training deep neural networks on large scientific data is a challenging task that requires enormous compute power, especially if no pre-trained models exist to initialize the process. We present a novel tournament method to train traditional as well as generative adversarial networks built on LBANN, a scalable deep learning framework optimized for HPC systems. LBANN combines multiple levels of parallelism and exploits some of the worlds largest supercomputers.We demonstrate our framework by creating a complex predictive model based on multi-variate data from high-energy-density physics containing hundreds of millions of images and hundreds of millions of scalar values derived from tens of millions of simulations of inertial confinement fusion. Our approach combines an HPC workflow and extends LBANN with optimized data ingestion and the new tournament-style training algorithm to produce a scalable neural network architecture using a CORAL-class supercomputer. Experimental results show that 64 trainers (1024 GPUs) achieve a speedup of 70.2× over a single trainer (16 GPUs) baseline, and an effective 109% parallel efficiency.

Motion planning, which is the problem of computing feasible paths in an environment for a movable object, has applications in many domains ranging from robotics, to intelligent CAD, to protein folding. The best methods for solving this PSPACE-hard problem are so-called sampling-based planners. Recent work introduced uniform spatial subdivision techniques for parallelizing sampling-based motion planning algorithms that scaled well. However, such methods are prone to load imbalance, as planning time depends on region characteristics and, for most problems, the heterogeneity of the sub problems increases as the number of processors increases. In this work, we introduce two techniques to address load imbalance in the parallelization of sampling-based motion planning algorithms: an adaptive work stealing approach and bulk-synchronous redistribution. We show that applying these techniques to representatives of the two major classes of parallel sampling-based motion planning algorithms, probabilistic roadmaps and rapidly-exploring random trees, results in a more scalable and load-balanced computation on more than 3,000 cores.

In this paper, we evaluate and compare the quality and structure of roadmaps constructed from parallelizing sampling-based motion planning algorithms against that of roadmaps constructed using sequential planner. Also, we make an argument and provide experimental results that show that motion planning problems involving heterogenous environments (common in most realistic and large-scale motion planning) is a natural fit for spatial subdivision-based parallel processing. Spatial subdivision-based parallel processing approach is suited for heterogeneous environments because it allows for local adaption in solving a global problem while taking advantage of scalability that is possible with parallel processing.

Rapidly-exploring Random Tree (RRT), like other sampling-based motion planning methods, has been very successful in solving motion planning problems. Even so, sampling-based planners cannot solve all problems of interest efficiently, so attention is increasingly turning to parallelizing them. However, one challenge in parallelizing RRT is the global computation and communication overhead of nearest neighbor search, a key operation in RRTs. This is a critical issue as it limits the scalability of previous algorithms. We present two parallel algorithms to address this problem. The first algorithm extends existing work by introducing a parameter that adjusts how much local computation is done before a global update. The second algorithm radially subdivides the configuration space into regions, constructs a portion of the tree in each region in parallel, and connects the subtrees,i removing cycles if they exist. By subdividing the space, we increase computation locality enabling a scalable result. We show that our approaches are scalable. We present results demonstrating almost linear scaling to hundreds of processors on a Linux cluster and a Cray XE6 machine.

Probabilistic Roadmap Methods (PRMs) are widely used motion planning methods that sample robot configurations (nodes) and connect them to form a graph (roadmap) containing feasible trajectories. Many PRM variants propose different strategies for each of the steps and choosing among them is problem dependent. Planning in heterogeneous environments and/or on parallel machines necessitates dividing the problem into regions where these choices have to be made for each one. Hand-selecting the best method for each region becomes infeasible. In particular, there are many ways to select connection candidates, and choosing the appropriate strategy is input dependent. In this paper, we present a general connection framework that adaptively selects a neighbor finding strategy from a candidate set of options. Our framework learns which strategy to use by examining their success rates and costs. It frees the user of the burden of selecting the best strategy and allows the selection to change over time. We perform experiments on rigid bodies of varying geometry and articulated linkages up to 37 degrees of freedom. Our results show that strategy performance is indeed problem/region dependent, and our adaptive method harnesses their strengths. Over all problems studied, our method differs the least from manual selection of the best method, and if one were to manually select a single method across all problems, the performance can be quite poor. Our method is able to adapt to changing sampling density and learns different strategies for each region when the problem is partitioned for parallelism.

Rapidly-Exploring Random Trees (RRTs) have been successful at finding feasible solutions for many types of problems. With motion planning becoming more computationally demanding, we turn to parallel motion planning for efficient solutions. Existing work on distributed RRTs has been limited by the overhead that global communication requires. A recent approach, Radial RRT, demonstrated a scalable algorithm that subdivides the space into regions to increase the computation locality. However, if an obstacle completely blocks RRT growth in a region, the planning space is not covered and is thus not probabilistically complete. We present a new algorithm, Blind RRT, which ignores obstacles during initial growth to efficiently explore the entire space. Because obstacles are ignored, free components of the tree become disconnected and fragmented. Blind RRT merges parts of the tree that have become disconnected from the root. We show how this algorithm can be applied to the Radial RRT framework allowing both scalability and effectiveness in motion planning. This method is a probabilistically complete approach to parallel RRTs. We show that our method not only scales but also overcomes the motion planning limitations that Radial RRT has in a series of difficult motion planning tasks.

This paper describes a scalable method for parallelizing sampling-based motion planning algorithms. It subdivides configuration space (C-space) into (possibly overlapping) regions and independently, in parallel, uses standard (sequential) sampling-based planners to construct roadmaps in each region. Next, in parallel, regional roadmaps in adjacent regions are connected to form a global roadmap. By subdividing the space and restricting the locality of connection attempts, we reduce the work and inter-processor communication associated with nearest neighbor calculation, a critical bottleneck for scalability in existing parallel motion planning methods. We show that our method is general enough to handle a variety of planning schemes, including the widely used Probabilistic Roadmap (PRM) and Rapidly-exploring Random Trees (RRT) algorithms. We compare our approach to two other existing parallel algorithms and demonstrate that our approach achieves better and more scalable performance. Our approach achieves almost linear scalability on a 2400 core LINUX cluster and on a 153,216 core Cray XE6 petascale machine.

Probabilistic Roadmap Methods (PRMs) are one of the most used classes of motion planning methods. These sampling-based methods generate robot configurations (nodes) and then connect them to form a graph (roadmap) containing representative feasible pathways. A key step in PRM roadmap construction involves identifying a set of candidate neighbors for each node. Traditionally, these candidates are chosen to be the k-closest nodes based on a given distance metric. In this paper, we propose a new neighbor selection policy called LocalRand(k,K'), that first computes the K' closest nodes to a specified node and then selects k of those nodes at random. Intuitively, LocalRand attempts to benefit from random sampling while maintaining the higher levels of local planner success inherent to selecting more local neighbors. We provide a methodology for selecting the parameters k and K' . We perform an experimental comparison which shows that for both rigid and articulated robots, LocalRand results in roadmaps that are better connected than the traditional k-closest policy or a purely random neighbor selection policy. The cost required to achieve these results is shown to be comparable to k-closest.