Batched sparse linear algebra operations in general, and solvers in particular, have become the major algorithmic development activity and foremost performance engineering effort in the numerical software libraries work on modern hardware with accelerators such as GPUs. Many applications, ECP and non-ECP alike, require simultaneous solutions of many small linear systems of equations that are structurally sparse in one form or another. In order to move towards high hardware utilization levels, it is important to provide these applications with appropriate interface designs to be both functionally efficient and performance portable and give full access to the appropriate batched sparse solvers running on modern hardware accelerators prevalent across DOE supercomputing sites since the inception of ECP. To this end, we present here a summary of recent advances on the interface designs in use by HPC software libraries supporting batched sparse linear algebra and the development of sparse batched kernel codes for solvers and preconditioners. We also address the potential interoperability opportunities to keep the corresponding software portable between the major hardware accelerators from AMD, Intel, and NVIDIA, while maintaining the appropriate disclosure levels conforming to the active NDA agreements. The presented interface specifications include a mix of batched band, sparse iterative, and sparse direct solvers with their accompanying functionality that is already required by the application codes or we anticipated to be needed in the near future. This report summarizes progress in Kokkos Kernels and the xSDK libraries MAGMA, Ginkgo, hypre, PETSc, and SuperLU.
Derivative computation is a key component of optimization, sensitivity analysis, uncertainty quantification, and the solving of nonlinear problems. Automatic differentiation (AD) is a powerful technique for evaluating such derivatives, and in recent years, has been integrated into programming environments such as Jax, PyTorch, and TensorFlow to support derivative computations needed for training of machine learning models, facilitating wide-spread use of these technologies. The C++ language has become the de facto standard for scientific computing due to numerous factors, yet language complexity has made the wide-spread adoption of AD technologies for C++ difficult, hampering the incorporation of powerful differentiable programming approaches into C++ scientific simulations. This is exacerbated by the increasing emergence of architectures, such as GPUs, with limited memory capabilities and requiring massive thread-level concurrency. C++ AD tools must effectively use these environments to bring novel scientific simulations to next-generation DOE experimental and observational facilities. In this project, we investigated source transformation-based automatic differentiation using LLVM compiler infrastructure to automatically generate portable and efficient gradient computations of Kokkos-based code. We have demonstrated that our proposed strategy is feasible by investigating the usage of a prototype LLVM-based source transformation tool to generate gradients of simple functions made of sequences of simple Kokkos parallel regions. Speedups of up to 500x compared to Sacado were observed on NVIDIA V100 GPU.
The properties of electrons in matter are of fundamental importance. They give rise to virtually all material properties and determine the physics at play in objects ranging from semiconductor devices to the interior of giant gas planets. Modeling and simulation of such diverse applications rely primarily on density functional theory (DFT), which has become the principal method for predicting the electronic structure of matter. While DFT calculations have proven to be very useful, their computational scaling limits them to small systems. We have developed a machine learning framework for predicting the electronic structure on any length scale. It shows up to three orders of magnitude speedup on systems where DFT is tractable and, more importantly, enables predictions on scales where DFT calculations are infeasible. Our work demonstrates how machine learning circumvents a long-standing computational bottleneck and advances materials science to frontiers intractable with any current solutions.
Solving large number of small linear systems is increasingly becoming a bottleneck in computational science applications. While dense linear solvers for such systems have been studied before, batched sparse linear solvers are just starting to emerge. In this paper, we discuss algorithms for solving batched sparse linear systems and their implementation in the Kokkos Kernels library. The new algorithms are performance portable and map well to the hierarchical parallelism available in modern accelerator architectures. The sparse matrix vector product (SPMV) kernel is the main performance bottleneck of the Krylov solvers we implement in this work. The implementation of the batched SPMV and its performance are therefore discussed thoroughly in this paper. The implemented kernels are tested on different Central Processing Unit (CPU) and Graphic Processing Unit (GPU) architectures. We also develop batched Conjugate Gradient (CG) and batched Generalized Minimum Residual (GMRES) solvers using the batched SPMV. Our proposed solver was able to solve 20,000 sparse linear systems on V100 GPUs with a mean speedup of 76x and 924x compared to using a parallel sparse solver with a block diagonal system with all the small linear systems, and compared to solving the small systems one at a time, respectively. We see mean speedup of 0.51 compared to dense batched solver of cuSOLVER on V100, while using lot less memory. Thorough performance evaluation on three different architectures and analysis of the performance are presented.
Solving large number of small linear systems is increasingly becoming a bottleneck in computational science applications. While dense linear solvers for such systems have been studied before, batched sparse linear solvers are just starting to emerge. In this paper, we discuss algorithms for solving batched sparse linear systems and their implementation in the Kokkos Kernels library. The new algorithms are performance portable and map well to the hierarchical parallelism available in modern accelerator architectures. The sparse matrix vector product (SPMV) kernel is the main performance bottleneck of the Krylov solvers we implement in this work. The implementation of the batched SPMV and its performance are therefore discussed thoroughly in this paper. The implemented kernels are tested on different Central Processing Unit (CPU) and Graphic Processing Unit (GPU) architectures. We also develop batched Conjugate Gradient (CG) and batched Generalized Minimum Residual (GMRES) solvers using the batched SPMV. Our proposed solver was able to solve 20,000 sparse linear systems on V100 GPUs with a mean speedup of 76x and 924x compared to using a parallel sparse solver with a block diagonal system with all the small linear systems, and compared to solving the small systems one at a time, respectively. We see mean speedup of 0.51 compared to dense batched solver of cuSOLVER on V100, while using lot less memory. Thorough performance evaluation on three different architectures and analysis of the performance are presented.
The generalized Dryja-Smith-Widlund (GDSW) preconditioner is a two-level overlapping Schwarz domain decomposition (DD) preconditioner that couples a classical one-level overlapping Schwarz preconditioner with an energy-minimizing coarse space. When used to accelerate the convergence rate of Krylov subspace iterative methods, the GDSW preconditioner provides robustness and scalability for the solution of sparse linear systems arising from the discretization of a wide range of partial different equations. In this paper, we present FROSch (Fast and Robust Schwarz), a domain decomposition solver package which implements GDSW-type preconditioners for both CPU and GPU clusters. To improve the solver performance on GPUs, we use a novel decomposition to run multiple MPI processes on each GPU, reducing both solver's computational and storage costs and potentially improving the convergence rate. This allowed us to obtain competitive or faster performance using GPUs compared to using CPUs alone. We demonstrate the performance of FROSch on the Summit supercomputer with NVIDIA V100 GPUs, where we used NVIDIA Multi-Process Service (MPS) to implement our decomposition strategy.The solver has a wide variety of algorithmic and implementation choices, which poses both opportunities and challenges for its GPU implementation. We conduct a thorough experimental study with different solver options including the exact or inexact solution of the local overlapping subdomain problems on a GPU. We also discuss the effect of using the iterative variant of the incomplete LU factorization and sparse-triangular solve as the approximate local solver, and using lower precision for computing the whole FROSch preconditioner. Overall, the solve time was reduced by factors of about 2× using GPUs, while the GPU acceleration of the numerical setup time depend on the solver options and the local matrix sizes.
The long-standing problem of predicting the electronic structure of matter on ultra-large scales (beyond 100,000 atoms) is solved with machine learning.
Artificial intelligence and machine learning (AI/ML) are becoming important tools for scientific modeling and simulation as in several other fields such as image analysis and natural language processing. ML techniques can leverage the computing power available in modern systems and reduce the human effort needed to configure experiments, interpret and visualize results, draw conclusions from huge quantities of raw data, and build surrogates for physics based models. Domain scientists in fields like fluid dynamics, microelectronics and chemistry can automate many of their most difficult and repetitive tasks or improve the design times by use of the faster ML-surrogates. However, modern ML and traditional scientific highperformance computing (HPC) tend to use completely different software ecosystems. While ML frameworks like PyTorch and TensorFlow provide Python APIs, most HPC applications and libraries are written in C++. Direct interoperability between the two languages is possible but is tedious and error-prone. In this work, we show that a compiler-based approach can bridge the gap between ML frameworks and scientific software with less developer effort and better efficiency. We use the MLIR (multi-level intermediate representation) ecosystem to compile a pre-trained convolutional neural network (CNN) in PyTorch to freestanding C++ source code in the Kokkos programming model. Kokkos is a programming model widely used in HPC to write portable, shared-memory parallel code that can natively target a variety of CPU and GPU architectures. Our compiler-generated source code can be directly integrated into any Kokkosbased application with no dependencies on Python or cross-language interfaces.
The focus of this project is to accelerate and transform the workflow of multiscale materials modeling by developing an integrated toolchain seamlessly combining DFT, SNAP, LAMMPS, (shown in Figure 1-1) and a machine-learning (ML) model that will more efficiently extract information from a smaller set of first-principles calculations. Our ML model enables us to accelerate first-principles data generation by interpolating existing high fidelity data, and extend the simulation scale by extrapolating high fidelity data (102 atoms) to the mesoscale (104 atoms). It encodes the underlying physics of atomic interactions on the microscopic scale by adapting a variety of ML techniques such as deep neural networks (DNNs), and graph neural networks (GNNs). We developed a new surrogate model for density functional theory using deep neural networks. The developed ML surrogate is demonstrated in a workflow to generate accurate band energies, total energies, and density of the 298K and 933K Aluminum systems. Furthermore, the models can be used to predict the quantities of interest for systems with more number of atoms than the training data set. We have demonstrated that the ML model can be used to compute the quantities of interest for systems with 100,000 Al atoms. When compared with 2000 Al system the new surrogate model is as accurate as DFT, but three orders of magnitude faster. We also explored optimal experimental design techniques to choose the training data and novel Graph Neural Networks to train on smaller data sets. These are promising methods that need to be explored in the future.
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.
A myriad of phenomena in materials science and chemistry rely on quantum-level simulations of the electronic structure in matter. While moving to larger length and time scales has been a pressing issue for decades, such large-scale electronic structure calculations are still challenging despite modern software approaches and advances in high-performance computing. The silver lining in this regard is the use of machine learning to accelerate electronic structure calculations – this line of research has recently gained growing attention. The grand challenge therein is finding a suitable machine-learning model during a process called hyperparameter optimization. This, however, causes a massive computational overhead in addition to that of data generation. We accelerate the construction of machine-learning surrogate models by roughly two orders of magnitude by circumventing excessive training during the hyperparameter optimization phase. We demonstrate our workflow for Kohn-Sham density functional theory, the most popular computational method in materials science and chemistry.
Triangle counting is a fundamental building block in graph algorithms. In this article, we propose a block-based triangle counting algorithm to reduce data movement during both sequential and parallel execution. Our block-based formulation makes the algorithm naturally suitable for heterogeneous architectures. The problem of partitioning the adjacency matrix of a graph is well-studied. Our task decomposition goes one step further: it partitions the set of triangles in the graph. By streaming these small tasks to compute resources, we can solve problems that do not fit on a device. We demonstrate the effectiveness of our approach by providing an implementation on a compute node with multiple sockets, cores and GPUs. The current state-of-the-art in triangle enumeration processes the Friendster graph in 2.1 seconds, not including data copy time between CPU and GPU. Using that metric, our approach is 20 percent faster. When copy times are included, our algorithm takes 3.2 seconds. This is 5.6 times faster than the fastest published CPU-only time.
Proceedings of PMBS 2022: Performance Modeling, Benchmarking and Simulation of High Performance Computer Systems, Held in conjunction with SC 2022: The International Conference for High Performance Computing, Networking, Storage and Analysis
We propose a new benchmark for high-performance (HP) computers. Similar to High Performance Conjugate Gradient (HPCG), the new benchmark is designed to rank computers based on how fast they can solve a sparse linear system of equations, exhibiting computational and communication requirements typical in many scientific applications. The main novelty of the new benchmark is that it is now based on Generalized Minimum Residual method (GMRES) (combined with Geometric Multi-Grid preconditioner and Gauss-Seidel smoother) and provides the flexibility to utilize lower precision arithmetic. This is motivated by new hardware architectures that deliver lower-precision arithmetic at higher performance. There are other machines that do not follow this trend. However, using a lower-precision arithmetic reduces the required amount of data transfer, which alone could improve solver performance. Considering these trends, an HP benchmark that allows the use of different precisions for solving important scientific problems will be valuable for many different disciplines, and we also hope to promote the design of future HP computers that can utilize mixed-precision arithmetic for achieving high application performance. We present our initial design of the new benchmark, its reference implementation, and the performance of the reference mixed (double and single) precision Geometric Multi-Grid solvers on current top-ranked architectures. We also discuss challenges of designing such a benchmark, along with our preliminary numerical results using 16-bit numerical values (half and bfloat precisions) for solving a sparse linear system of equations.
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 generalized Dryja-Smith-Widlund (GDSW)-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 message passing interface (MPI)- parallel implementation of multilevel 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 the best of our knowledge, this is the first time two-level Schwarz preconditioners have been applied to the ice sheet problem and a scalable preconditioner has been used for the coupled problem. The preconditioner 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 subdomains. 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 nonuniform meshes for the Greenland ice sheet. 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.
Given a graph, finding the distance-2 maximal independent set (MIS-2) of the vertices is a problem that is useful in several contexts such as algebraic multigrid coarsening or multilevel graph partitioning. Such multilevel methods rely on finding the independent vertices so they can be used as seeds for aggregation in a multilevel scheme. We present a parallel MIS-2 algorithm to improve performance on modern accelerator hardware. This algorithm is implemented using the Kokkos programming model to enable performance portability. We demonstrate the portability of the algorithm and the performance on a variety of architectures (x86/ARM CPUs and NVIDIA/AMD GPUs). The resulting algorithm is also deterministic, producing an identical result for a given input across all of these platforms. The new MIS-2 implementation outperforms implementations in state of the art libraries like CUSP and ViennaCL by 3-8x while producing similar quality results. We further demonstrate the benefits of this approach by developing parallel graph coarsening scheme for two different use cases. First, we develop an algebraic multigrid (AMG) aggregation scheme using parallel MIS-2 and demonstrate the benefits as opposed to previous approaches used in the MueLu multigrid package in Trilinos. We also describe an approach for implementing a parallel multicolor 'cluster' Gauss-Seidel preconditioner using this MIS-2 coarsening, and demonstrate better performance with an efficient, parallel, mul-ticolor Gauss-Seidel algorithm.
To keep pace with the demand for innovation through scientific computing, modern scientific software development is increasingly reliant upon a rich and diverse ecosystem of software libraries and toolchains. Research software engineers (RSEs) responsible for that infrastructure perform highly integrative work, acting as a bridge between the hardware, the needs of researchers, and the software layers situated between them; relatively little, however, has been written about the role played by RSEs in that work and what support they need to thrive. To that end, we present a two-part report on the development of half-precision floating point support in the Kokkos Ecosystem. Half-precision computation is a promising strategy for increasing performance in numerical computing and is particularly attractive for emerging application areas (e.g., machine learning), but developing practicable, portable, and user-friendly abstractions is a nontrivial task. In the first half of the paper, we conduct an engineering study on the technical implementation of the Kokkos half-precision scalar feature and showcase experimental results; in the second half, we offer an experience report on the challenges and lessons learned during feature development by the first author. We hope our study provides a holistic view on scientific library development and surfaces opportunities for future studies into effective strategies for RSEs engaged in such work.
Rapid growth in data, computational methods, and computing power is driving a remarkable revolution in what variously is termed machine learning (ML), statistical learning, computational learning, and artificial intelligence. In addition to highly visible successes in machine-based natural language translation, playing the game Go, and self-driving cars, these new technologies also have profound implications for computational and experimental science and engineering, as well as for the exascale computing systems that the Department of Energy (DOE) is developing to support those disciplines. Not only do these learning technologies open up exciting opportunities for scientific discovery on exascale systems, they also appear poised to have important implications for the design and use of exascale computers themselves, including high-performance computing (HPC) for ML and ML for HPC. The overarching goal of the ExaLearn co-design project is to provide exascale ML software for use by Exascale Computing Project (ECP) applications, other ECP co-design centers, and DOE experimental facilities and leadership class computing facilities.
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.
