This report includes a compilation of several slide presentations: 1) Interatomic Potentials for Materials Science and Beyond–Advances in Machine Learned Spectral Neighborhood Analysis Potentials (Wood); 2) Agile Materials Science and Advanced Manufacturing through AI/ML (de Oca Zapiain); 3) Machine Learning for DFT Calculations (Rajamanickam); 4) Structure-preserving ML discovery of a quantum-to-continuum codesign stack (Trask); and 5) IBM Overview of Accelerated Discovery Technology (Pitera)
State-of-the-art engineering and science codes have grown in complexity dramatically over the last two decades. Application teams have adopted more sophisticated development strategies, leveraging third party libraries, deploying comprehensive testing, and using advanced debugging and profiling tools. In today's environment of diverse hardware platforms, these applications also desire performance portability-avoiding the need to duplicate work for various platforms. The Kokkos EcoSystem provides that portable software stack. Based on the Kokkos Core Programming Model, the EcoSystem provides math libraries, interoperability capabilities with Python and Fortran, and Tools for analyzing, debugging, and optimizing applications. In this article, we overview the components, discuss some specific use cases, and highlight how codesigning these components enables a more developer friendly experience.
Garg, Raveesh; Qin, Eric; Martinez, Francisco M.; Guirado, Robert; Jain, Akshay; Abadal, Sergi; Abellan, Jose L.; Acacio, Manuel E.; Alarcon, Eduard; Rajamanickam, Sivasankaran; Krishna, Tushar
Graph Neural Networks (GNNs) have garnered a lot of recent interest because of their success in learning representations from graph-structured data across several critical applications in cloud and HPC. Owing to their unique compute and memory characteristics that come from an interplay between dense and sparse phases of computations, the emergence of reconfigurable dataflow (aka spatial) accelerators offers promise for acceleration by mapping optimized dataflows (i.e., computation order and parallelism) for both phases. The goal of this work is to characterize and understand the design-space of dataflow choices for running GNNs on spatial accelerators in order for the compilers to optimize the dataflow based on the workload. Specifically, we propose a taxonomy to describe all possible choices for mapping the dense and sparse phases of GNNs spatially and temporally over a spatial accelerator, capturing both the intra-phase dataflow and the inter-phase (pipelined) dataflow. Using this taxonomy, we do deep-dives into the cost and benefits of several dataflows and perform case studies on implications of hardware parameters for dataflows and value of flexibility to support pipelined execution.
Graph partitioning has been an important tool to partition the work among several processors to minimize the communication cost and balance the workload. While accelerator-based supercomputers are emerging to be the standard, the use of graph partitioning becomes even more important as applications are rapidly moving to these architectures. However, there is no distributed-memory-parallel, multi-GPU graph partitioner available for applications. We developed a spectral graph partitioner, Sphynx, using the portable, accelerator-friendly stack of the Trilinos framework. In Sphynx, we allow using different preconditioners and exploit their unique advantages. We use Sphynx to systematically evaluate the various algorithmic choices in spectral partitioning with a focus on the GPU performance. We perform those evaluations on two distinct classes of graphs: regular (such as meshes, matrices from finite element methods) and irregular (such as social networks and web graphs), and show that different settings and preconditioners are needed for these graph classes. The experimental results on the Summit supercomputer show that Sphynx is the fastest alternative on irregular graphs in an application-friendly setting and obtains a partitioning quality close to ParMETIS on regular graphs. When compared to nvGRAPH on a single GPU, Sphynx is faster and obtains better balance and better quality partitions. Sphynx provides a good and robust partitioning method across a wide range of graphs for applications looking for a GPU-based partitioner.
Abdelfattah, Ahmad; Anzt, Hartwig; Ayala, Alan; Boman, Erik G.; Carson, Erin C.; Cayrols, Sebastien; Cojean, Terry; Dongarra, Jack J.; Falgout, Rob; Gates, Mark; G, R\{U}Tzmacher; Higham, Nicholas J.; Kruger, Scott E.; Li, Sherry; Lindquist, Neil; Liu, Yang; Loe, Jennifer A.; Nayak, Pratik; Osei-Kuffuor, Daniel; Pranesh, Sri; Rajamanickam, Sivasankaran; Ribizel, Tobias; Smith, Bryce; Swirydowicz, Kasia; Thomas, Stephen J.; Tomov, Stanimire; Tsai, Yaohung M.; Yamazaki, Ichitaro; Yang, Urike M.
Over the last year, the ECP xSDK-multiprecision effort has made tremendous progress in developing and deploying new mixed precision technology and customizing the algorithms for the hardware deployed in the ECP flagship supercomputers. The effort also has succeeded in creating a cross-laboratory community of scientists interested in mixed precision technology and now working together in deploying this technology for ECP applications. In this report, we highlight some of the most promising and impactful achievements of the last year. Among the highlights we present are: Mixed precision IR using a dense LU factorization and achieving a 1.8× speedup on Spock; results and strategies for mixed precision IR using a sparse LU factorization; a mixed precision eigenvalue solver; Mixed Precision GMRES-IR being deployed in Trilinos, and achieving a speedup of 1.4× over standard GMRES; compressed Basis (CB) GMRES being deployed in Ginkgo and achieving an average 1.4× speedup over standard GMRES; preparing hypre for mixed precision execution; mixed precision sparse approximate inverse preconditioners achieving an average speedup of 1.2×; and detailed description of the memory accessor separating the arithmetic precision from the memory precision, and enabling memory-bound low precision BLAS 1/2 operations to increase the accuracy by using high precision in the computations without degrading the performance. We emphasize that many of the highlights presented here have also been submitted to peer-reviewed journals or established conferences, and are under peer-review or have already been published.
As the push towards exascale hardware has increased the diversity of system architectures, performance portability has become a critical aspect for scientific software. We describe the Kokkos Performance Portable Programming Model that allows developers to write single source applications for diverse high performance computing architectures. Kokkos provides key abstractions for both the compute and memory hierarchy of modern hardware. Here, we describe the novel abstractions that have been added to Kokkos recently such as hierarchical parallelism, containers, task graphs, and arbitrary-sized atomic operations. We demonstrate the performance of these new features with reproducible benchmarks on CPUs and GPUs.
There is a growing interest in custom spatial accelerators for machine learning applications. These accelerators employ a spatial array of processing elements (PEs) interacting via custom buffer hierarchies and networks-on-chip. The efficiency of these accelerators comes from employing optimized dataflow (i.e., spatial/temporal partitioning of data across the PEs and fine-grained scheduling) strategies to optimize data reuse. The focus of this work is to evaluate these accelerator architectures using a tiled general matrix-matrix multiplication (GEMM) kernel. To do so, we develop a framework that finds optimized mappings (dataflow and tile sizes) for a tiled GEMM for a given spatial accelerator and workload combination, leveraging an analytical cost model for runtime and energy. Our evaluations over five spatial accelerators demonstrate that the tiled GEMM mappings systematically generated by our framework achieve high performance on various GEMM workloads and accelerators.
Support for lower precision computation is becoming more common in accelerator hardware due to lower power usage, reduced data movement and increased computational performance. However, computational science and engineering (CSE) problems require double precision accuracy in several domains. This conflict between hardware trends and application needs has resulted in a need for multiprecision strategies at the linear algebra algorithms level if we want to exploit the hardware to its full potential while meeting the accuracy requirements. In this paper, we focus on preconditioned sparse iterative linear solvers, a key kernel in several CSE applications. We present a study of multiprecision strategies for accelerating this kernel on GPUs. We seek the best methods for incorporating multiple precisions into the GMRES linear solver; these include iterative refinement and parallelizable preconditioners. Our work presents strategies to determine when multiprecision GMRES will be effective and to choose parameters for a multiprecision iterative refinement solver to achieve better performance. We use an implementation that is based on the Trilinos library and employs Kokkos Kernels for performance portability of linear algebra kernels. Performance results demonstrate the promise of multiprecision approaches and demonstrate even further improvements are possible by optimizing low-level kernels.
We present a numerical modeling workflow based on machine learning (ML) which reproduces the total energies produced by Kohn-Sham density functional theory (DFT) at finite electronic temperature to within chemical accuracy at negligible computational cost. Based on deep neural networks, our workflow yields the local density of states (LDOS) for a given atomic configuration. From the LDOS, spatially-resolved, energy-resolved, and integrated quantities can be calculated, including the DFT total free energy, which serves as the Born-Oppenheimer potential energy surface for the atoms. We demonstrate the efficacy of this approach for both solid and liquid metals and compare results between independent and unified machine-learning models for solid and liquid aluminum. Our machine-learning density functional theory framework opens up the path towards multiscale materials modeling for matter under ambient and extreme conditions at a computational scale and cost that is unattainable with current algorithms.
Sparsity, which occurs in both scientific applications and Deep Learning (DL) models, has been a key target of optimization within recent ASIC accelerators due to the potential memory and compute savings. These applications use data stored in a variety of compression formats. We demonstrate that both the compactness of different compression formats and the compute efficiency of the algorithms enabled by them vary across tensor dimensions and amount of sparsity. Since DL and scientific workloads span across all sparsity regions, there can be numerous format combinations for optimizing memory and compute efficiency. Unfortunately, many proposed accelerators operate on one or two fixed format combinations. This work proposes hardware extensions to accelerators for supporting numerous format combinations seamlessly and demonstrates ∼ 4 × speedup over performing format conversions in software.
Learning 3D representations that generalize well to arbitrarily oriented inputs is a challenge of practical importance in applications varying from computer vision to physics and chemistry. We propose a novel multi-resolution convolutional architecture for learning over concentric spherical feature maps, of which the single sphere representation is a special case. Our hierarchical architecture is based on alternatively learning to incorporate both intra-sphere and inter-sphere information. We show the applicability of our method for two different types of 3D inputs, mesh objects, which can be regularly sampled, and point clouds, which are irregularly distributed. We also propose an efficient mapping of point clouds to concentric spherical images, thereby bridging spherical convolutions on grids with general point clouds. We demonstrate the effectiveness of our approach in improving state-of-the-art performance on 3D classification tasks with rotated data.
Garg, Raveesh; Qin, Eric; Martinez, Francisco M.; Guirado, Robert; Jain, Akshay; Abadal, Sergi; Abellan, Jose L.; Acacio, Manuel E.; Alarcon, Eduard; Rajamanickam, Sivasankaran; Krishna, Tushar
Recently, Graph Neural Networks (GNNs) have received a lot of interest because of their success in learning representations from graph structured data. However, GNNs exhibit different compute and memory characteristics compared to traditional Deep Neural Networks (DNNs). Graph convolutions require feature aggregations from neighboring nodes (known as the aggregation phase), which leads to highly irregular data accesses. GNNs also have a very regular compute phase that can be broken down to matrix multiplications (known as the combination phase). All recently proposed GNN accelerators utilize different dataflows and microarchitecture optimizations for these two phases. Different communication strategies between the two phases have been also used. However, as more custom GNN accelerators are proposed, the harder it is to qualitatively classify them and quantitatively contrast them. In this work, we present a taxonomy to describe several diverse dataflows for running GNN inference on accelerators. This provides a structured way to describe and compare the design-space of GNN accelerators.
MLIR (Multi-Level Intermediate Representation), is an extensible compiler framework that supports high-level data structures and operation constructs. These higher-level code representations are particularly applicable to the artificial intelligence and machine learning (AI/ML) domain, allowing developers to more easily support upcoming heterogeneous AI/ML accelerators and develop flexible domain specific compilers/frameworks with higher-level intermediate representations (IRs) and advanced compiler optimizations. The result of using MLIR within the LLVM compiler framework is expected to yield significant improvement in the quality of generated machine code, which in turn will result in improved performance and hardware efficiency
Solving dense systems of linear equations is essential in applications encountered in physics, mathematics, and engineering. This paper describes our current efforts toward the development of the ADELUS package for current and next generation distributed, accelerator-based, high-performance computing platforms. The package solves dense linear systems using partial pivoting LU factorization on distributed-memory systems with CPUs/GPUs. The matrix is block-mapped onto distributed memory on CPUs/GPUs and is solved as if it was torus-wrapped for an optimal balance of computation and communication. A permutation operation is performed to restore the results so the torus-wrap distribution is transparent to the user. This package targets performance portability by leveraging the abstractions provided in the Kokkos and Kokkos Kernels libraries. Comparison of the performance gains versus the state-of-the-art SLATE and DPLASMA GESV functionalities on the Summit supercomputer are provided. Preliminary performance results from large-scale electromagnetic simulations using ADELUS are also presented. The solver achieves 7.7 Petaflops on 7600 GPUs of the Sierra supercomputer translating to 16.9% efficiency.
Solving dense systems of linear equations is essential in applications encountered in physics, mathematics, and engineering. This paper describes our current efforts toward the development of the ADELUS package for current and next generation distributed, accelerator-based, high-performance computing platforms. The package solves dense linear systems using partial pivoting LU factorization on distributed-memory systems with CPUs/GPUs. The matrix is block-mapped onto distributed memory on CPUs/GPUs and is solved as if it was torus-wrapped for an optimal balance of computation and communication. A permutation operation is performed to restore the results so the torus-wrap distribution is transparent to the user. This package targets performance portability by leveraging the abstractions provided in the Kokkos and Kokkos Kernels libraries. Comparison of the performance gains versus the state-of-the-art SLATE and DPLASMA GESV functionalities on the Summit supercomputer are provided. Preliminary performance results from large-scale electromagnetic simulations using ADELUS are also presented. The solver achieves 7.7 Petaflops on 7600 GPUs of the Sierra supercomputer translating to 16.9% efficiency.
To meet the extreme compute demands for deep learning across commercial and scientific applications, dataflow accelerators are becoming increasingly popular. While these “domain-specific” accelerators are not fully programmable like CPUs and GPUs, they retain varying levels of flexibility with respect to data orchestration, i.e., dataflow and tiling optimizations to enhance efficiency. There are several challenges when designing new algorithms and mapping approaches to execute the algorithms for a target problem on new hardware. Previous works have addressed these challenges individually. To address this challenge as a whole, in this work, we present a HW-SW codesign ecosystem for spatial accelerators called Union within the popular MLIR compiler infrastructure. Our framework allows exploring different algorithms and their mappings on several accelerator cost models. Union also includes a plug-and-play library of accelerator cost models and mappers which can easily be extended. The algorithms and accelerator cost models are connected via a novel mapping abstraction that captures the map space of spatial accelerators which can be systematically pruned based on constraints from the hardware, workload, and mapper. We demonstrate the value of Union for the community with several case studies which examine offloading different tensor operations (CONV/GEMM/Tensor Contraction) on diverse accelerator architectures using different mapping schemes.
To meet the extreme compute demands for deep learning across commercial and scientific applications, dataflow accelerators are becoming increasingly popular. While these “domain-specific” accelerators are not fully programmable like CPUs and GPUs, they retain varying levels of flexibility with respect to data orchestration, i.e., dataflow and tiling optimizations to enhance efficiency. There are several challenges when designing new algorithms and mapping approaches to execute the algorithms for a target problem on new hardware. Previous works have addressed these challenges individually. To address this challenge as a whole, in this work, we present a HW-SW codesign ecosystem for spatial accelerators called Union within the popular MLIR compiler infrastructure. Our framework allows exploring different algorithms and their mappings on several accelerator cost models. Union also includes a plug-and-play library of accelerator cost models and mappers which can easily be extended. The algorithms and accelerator cost models are connected via a novel mapping abstraction that captures the map space of spatial accelerators which can be systematically pruned based on constraints from the hardware, workload, and mapper. We demonstrate the value of Union for the community with several case studies which examine offloading different tensor operations (CONV/GEMM/Tensor Contraction) on diverse accelerator architectures using different mapping schemes.
Numerical simulations of Greenland and Antarctic ice sheets involve the solution of large-scale highly nonlinear systems of equations on complex shallow geometries. This work is concerned with the construction of Schwarz preconditioners for the solution of the associated tangent problems, which are challenging for solvers mainly because of the strong anisotropy of the meshes and wildly changing boundary conditions that can lead to poorly constrained problems on large portions of the domain. Here, two-level GDSW (Generalized Dryja–Smith–Widlund) type Schwarz preconditioners are applied to different land ice problems, i.e., a velocity problem, a temperature problem, as well as the coupling of the former two problems. We employ the MPI-parallel implementation of multi-level Schwarz preconditioners provided by the package FROSch (Fast and Robust Schwarz)from the Trilinos library. The strength of the proposed preconditioner is that it yields out-of-the-box scalable and robust preconditioners for the single physics problems. To our knowledge, this is the first time two-level Schwarz preconditioners are applied to the ice sheet problem and a scalable preconditioner has been used for the coupled problem. The pre-conditioner for the coupled problem differs from previous monolithic GDSW preconditioners in the sense that decoupled extension operators are used to compute the values in the interior of the sub-domains. Several approaches for improving the performance, such as reuse strategies and shared memory OpenMP parallelization, are explored as well. In our numerical study we target both uniform meshes of varying resolution for the Antarctic ice sheet as well as non uniform meshes for the Greenland ice sheet are considered. We present several weak and strong scaling studies confirming the robustness of the approach and the parallel scalability of the FROSch implementation. Among the highlights of the numerical results are a weak scaling study for up to 32 K processor cores (8 K MPI-ranks and 4 OpenMP threads) and 566 M degrees of freedom for the velocity problem as well as a strong scaling study for up to 4 K processor cores (and MPI-ranks) and 68 M degrees of freedom for the coupled problem.
Parallel implementations of linear iterative solvers generally alternate between phases of data exchange and phases of local computation. Increasingly large problem sizes and more heterogeneous compute architectures make load balancing and the design of low latency network interconnects that are able to satisfy the communication requirements of linear solvers very challenging tasks. In particular, global communication patterns such as inner products become increasingly limiting at scale. We explore the use of asynchronous communication based on one-sided Message Passing Interface primitives in the context of domain decomposition solvers. In particular, a scalable asynchronous two-level Schwarz method is presented. We discuss practical issues encountered in the development of a scalable solver and show experimental results obtained on a state-of-the-art supercomputer system that illustrate the benefits of asynchronous solvers in load balanced as well as load imbalanced scenarios. Using the novel method, we can observe speedups of up to four times over its classical synchronous equivalent.