Accelerating the SVD bi-diagonalization of a batch of small matrices using GPUs
Not Available
Engineering topics
Publications and source records attributed to Tomov, Stanimire.
Not Available
The emergence of power efficiency as a primary constraint in processor and system design poses new challenges concerning power and energy awareness for numerical libraries and scientific applications. Power consumption also plays a major role in the design of data centers, which may house petascale or exascale-level computing systems. At these extreme scales, understanding and improving the energy efficiency of numerical libraries and their related applications becomes a crucial part of the successful implementation and operation of the computing system. In this paper, we study and investigate the practice of controlling a compute system's power usage, and we explore how different power caps affect the performance of numerical algorithms with different computational intensities. Further, we determine the impact, in terms of performance and energy usage, that these caps have on a system running scientific applications. This analysis will enable us to characterize the types of algorithms that benefit most from these power management schemes. Our experiments are performed using a set of representative kernels and several popular scientific benchmarks. Lastly, we quantify a number of power and performance measurements and draw observations and conclusions that can be viewed as a roadmap to achieving energy efficiency in the design and execution of scientific algorithms.
Key components of CEED software development involve fast finite element operator storage and evaluation, architecture optimizations, performant algorithms for all orders, global kernels for finite element operators, efficient use of the memory sub-system, optimal data locality and motion, enhanced scalability and parallelism, and fast tensor contractions. As ultimate goals of the project, CEED is exploring and identifying the best algorithms for the full range of discretizations and applying algorithmic and software development to support ECP applications’ needs. In this milestone we report on the initial integration of CEED software in the selected first-wave ECP/CEED apps and the continued exploration of other ECP applications as second-wave (milestone CEED-MS23) and third-wave (milestone CEED-MS35) ECP/CEED application candidates. An extensive evaluation of various approaches for GPU acceleration of the CEED computational motives is also a major focus of the report, both in the settings of CEED’s Nekbone and Laghos miniapps, as well as versions of CEED’s bake-off problems (introduced in milestone CEED-MS6). The first-wave ECP applications targeted for coupling with CEED software include the ExaSMR and MARBL applications. We also report on near-term collaboration with the seed-funded Urban Systems project, as well as engagements with ACME, ExaAM, GEOS and the Application Assessment and Proxy App projects. The ExaSMR project involves coupled thermal-hydraulics/neutronics, with the latter based on Monte Carlo methods and the former based on high-order discretizations of the Navier-Stokes and energy equations for accurate simulation of turbulent transport. The project requires fast and efficient turbulence simulation capabilities that will scale to exascale platforms. Work within CEED will ensure that all aspects of the Nek5000 workflow scale to the target problem space, which involves > 10 10 degrees of freedom in typical production runs. CEED’s goal is to provide implementations that perform with maximum attainable efficiency and deliver low turn-around times. A deliverable for this milestone is beta-release of the GPU-based variant of Nek5000. MARBL is targeting high-energy-density physics problems, including single fluid multi-material hydrodynamics and radiation/magnetic diffusion simulation, with applications in inertial confined fusion, pulsed power experiments, and equation of state/material strength analysis. The code uses high-order methods based arbitrary Lagrangian-Eulerian, direct Eulerian, and unstructured adaptive mesh refinement. A deliverable for this milestone is collaboration with MARBL for improved performance based on partially-assembled operators. The Urban exascale project is planning to couple real-time data with high performance simulations to study particulate transport in cities. Nek5000 will be used for large eddy and RANS (Reynolds-averaged Navier-Stokes) simulations of turbulent transport in these domains. Areas where the CEED team is aiding the Urban Canyon group include creating high-order meshes for urban geometries and developing boundary conditions appropriate for simulations close to real-life urban conditions. The software artifacts delivered as part of this milestone include GPU-enabled versions of the Nekbone and Laghos miniapps, which are provided through the CEED website, http://ceed.exascaleproject.org and the CEED GitHub organization, http://github.com/ceed. In addition to details and results from the above R&D efforts, in this document we are also reporting on other project-wide activities performed in Q4 of FY17, including: the CEED first annual meeting and participation of CEED researchers in a variety of outreach activities.
The CEED co-design center is developing a number of kernels, bake-off / benchmark problems (BPs) and mini-applications (miniapps) that capture the unique requirements of high-order finite element algorithms and use them to influence standards, best practices and vendors to generate more portable, better performing code across diverse hardware designs. This is a main activity in CEED’s Hardware, Applications and Finite Element thrusts. CEED’s miniapps, a.k.a. CEEDlings, are small, standalone applications - surrogates for sub-components of interest from the selected first wave of applications (the MARBL/LLNLApp and ExaSMR projects). The BPs are even smaller standalone programs that solve a simple physical problem, including discretization, linear and/or nonlinear solve, and potentially a relatively small number of time steps. A BP may be a proxy/surrogate for part of a miniapp or another problem of general interest. In CEED, both miniapps and BPs will be used extensively in the project to evaluate and compare the performance of algorithms in specific contexts. The parts of the miniapps and BPs that play an important role in their overall performance will be identified and separated as standalone kernels. These kernels and their interaction inside BPs, miniapps and the actual applications, will be the main focus of our optimization efforts. All three standalone pieces of software (miniapps, BPs, and kernels) will serve as a basis for engaging and collaborating with hardware vendors and software technologies (ST) projects. These partners can start at the bottom of the hierarchy, e.g. with the kernels, and easily transition to BPs and miniapps as all three families will be produced in the same CEED environment. In this milestone we identified the initial kernels, bake-off problems and miniapps and delivered software and documentation for them through the new CEED website, http://ceed.exascaleproject.org. We specifically identified the Nekbone miniapp for the ExaSMR application and developed a new miniapp, Laghos, for the MARBL/LLNLApp application. We also formulated four initial benchmark problems (BP1- BP4) with clearly defined performance metrics and performed an initial bake-off comparison between the Nek5000 (spectral element) and MFEM (high-order finite element) technologies. As part of the milestone, we also engaged vendors (AMD) and ST projects (MPICH, STRUMPACK). In this document we are reporting some details and results from these R&D activities, as well as additional project-wide activities performed in Q3 of FY17, including: making CEED software available on GitHub, developing the project website, and the preparation of CEED’s first annual meeting.
In this paper we explore the performance portability of directives provided by OpenMP 4 and OpenACC to program various types of node architectures with attached accelerators, both self-hosted multicore and offload multicore/GPU. Our goal is to examine how successful OpenACC and the newer offload features of OpenMP 4.5 are for moving codes between architectures, how much tuning might be required and what lessons we can learn from this experience. To do this, we use examples of algorithms with varying computational intensities for our evaluation, as both compute and data access efficiency are important considerations for overall application performance. We implement these kernels using various methods provided by newer OpenACC and OpenMP implementations, and we evaluate their performance on various platforms including both X86_64 with attached NVIDIA GPUs, self-hosted Intel Xeon Phi KNL, as well as an X86_64 host system with Intel Xeon Phi coprocessors. In this paper, we explain what factors affected the performance portability such as how to pick the right programming model, its programming style, its availability on different platforms, and how well compilers can optimize and target to multiple platforms.
In this paper we unveil some performance and energy efficiency frontiers for sparse computations on GPU-based supercomputers. We compare the resource efficiency of different sparse matrix–vector products (SpMV) taken from libraries such as cuSPARSE and MAGMA for GPU and Intel’s MKL for multicore CPUs, and develop a GPU sparse matrix–matrix product (SpMM) implementation that handles the simultaneous multiplication of a sparse matrix with a set of vectors in block-wise fashion. While a typical sparse computation such as the SpMV reaches only a fraction of the peak of current GPUs, we show that the SpMM succeeds in exceeding the memory-bound limitations of the SpMV. We integrate this kernel into a GPU-accelerated Locally Optimal Block Preconditioned Conjugate Gradient (LOBPCG) eigensolver. LOBPCG is chosen as a benchmark algorithm for this study as it combines an interesting mix of sparse and dense linear algebra operations that is typical for complex simulation applications, and allows for hardware-aware optimizations. In a detailed analysis we compare the performance and energy efficiency against a multi-threaded CPU counterpart. The reported performance and energy efficiency results are indicative of sparse computations on supercomputers.
Abstract—We demonstrate the systematic implementation of recently-developed fast explicit kinetic integration algorithms that solve efficiently N coupled ordinary differential equations (subject to initial conditions) on modern GPUs. We take representative test cases (Type Ia supernova explosions) and demonstrate two or more orders of magnitude increase in efficiency for solving such systems (of realistic thermonuclear networks coupled to fluid dynamics). This implies that important coupled, multiphysics problems in various scientific and technical disciplines that were intractable, or could be simulated only with highly schematic kinetic networks, are now computationally feasible. As examples of such applications we present the computational techniques developed for our ongoing deployment of these new methods on modern GPU accelerators. We show that similarly to many other scientific applications, ranging from national security to medical advances, the computation can be split into many independent computational tasks, each of relatively small-size. As the size of each individual task does not provide sufficient parallelism for the underlying hardware, especially for accelerators, these tasks must be computed concurrently as a single routine, that we call batched routine, in order to saturate the hardware with enough work.
Krylov subspace iterative solvers are often the method of choice when solving large sparse linear systems. At the same time, hardware accelerators such as graphics processing units continue to offer significant floating point performance gains for matrix and vector computations through easy-to-use libraries of computational kernels. However, as these libraries are usually composed of a well optimized but limited set of linear algebra operations, applications that use them often fail to reduce certain data communications, and hence fail to leverage the full potential of the accelerator. In this study, we target the acceleration of Krylov subspace iterative methods for graphics processing units, and in particular the Biconjugate Gradient Stabilized solver that significant improvement can be achieved by reformulating the method to reduce data-communications through application-specific kernels instead of using the generic BLAS kernels, e.g. as provided by NVIDIA’s cuBLAS library, and by designing a graphics processing unit specific sparse matrix-vector product kernel that is able to more efficiently use the graphics processing unit’s computing power. Furthermore, we derive a model estimating the performance improvement, and use experimental data to validate the expected runtime savings. Finally, considering that the derived implementation achieves significantly higher performance, we assert that similar optimizations addressing algorithm structure, as well as sparse matrix-vector, are crucial for the subsequent development of high-performance graphics processing units accelerated Krylov subspace iterative methods.
Abstract not provided.
Scientific applications require solvers that work on many small size problems that are independent from each other. At the same time, the high-end hardware evolves rapidly and becomes ever more throughput-oriented and thus there is an increasing need for an effective approach to develop energy-efficient, high-performance codes for these small matrix problems that we call batched factorizations. The many applications that need this functionality could especially benefit from the use of GPUs, which currently are four to five times more energy efficient than multicore CPUs on important scientific workloads. This study, consequently, describes the development of the most common, one-sided factorizations, Cholesky, LU, and QR, for a set of small dense matrices. The algorithms we present together with their implementations are, by design, inherently parallel. In particular, our approach is based on representing the process as a sequence of batched BLAS routines that are executed entirely on a GPU. Importantly, this is unlike the LAPACK and the hybrid MAGMA factorization algorithms that work under drastically different assumptions of hardware design and efficiency of execution of the various computational kernels involved in the implementation. Thus, our approach is more efficient than what works for a combination of multicore CPUs and GPUs for the problems sizes of interest of the application use cases. The paradigm where upon a single chip (a GPU or a CPU) factorizes a single problem at a time is not at all efficient in our applications’ context. We illustrate all of these claims through a detailed performance analysis. With the help of profiling and tracing tools, we guide our development of batched factorizations to achieve up to two-fold speedup and three-fold better energy efficiency as compared against our highly optimized batched CPU implementations based on MKL library. Finally, the tested system featured two sockets of Intel Sandy Bridge CPUs and we compared with a batched LU factorizations featured in the CUBLAS library for GPUs, we achieve as high as 2.5× speedup on the NVIDIA K40 GPU.
The manycore paradigm shift, and the resulting change in modern computer architectures, has made the development of optimal numerical routines extremely challenging. In this work, we target the development of numerical algorithms and implementations for Xeon Phi coprocessor architecture designs. In particular, we examine and optimize the general and symmetric matrix-vector multiplication routines (gemv/symv), which are some of the most heavily used linear algebra kernels in many important engineering and physics applications. We describe a successful approach on how to address the challenges for this problem, starting with our algorithm design, performance analysis and programing model and moving to kernel optimization. Our goal, by targeting low-level and easy to understand fundamental kernels, is to develop new optimization strategies that can be effective elsewhere for use on manycore coprocessors, and to show significant performance improvements compared to existing state-of-the-art implementations. Therefore, in addition to the new optimization strategies, analysis, and optimal performance results, we finally present the significance of using these routines/strategies to accelerate higher-level numerical algorithms for the eigenvalue problem (EVP) and the singular value decomposition (SVD) that by themselves are foundational for many important applications.
As hardware evolves, an increasingly effective approach to develop energy efficient, high-performance solvers, is to design them to work on many small and independent problems. Indeed, many applications already need this functionality, especially for GPUs, which are known to be currently about four to five times more energy efficient than multicore CPUs for every floating-point operation. In this paper, we describe the development of the main one-sided factorizations: LU, QR, and Cholesky; that are needed for a set of small dense matrices to work in parallel. We refer to such algorithms as batched factorizations. Our approach is based on representing the algorithms as a sequence of batched BLAS routines for GPU-contained execution. Note that this is similar in functionality to the LAPACK and the hybrid MAGMA algorithms for large-matrix factorizations. But it is different from a straightforward approach, whereby each of GPU's symmetric multiprocessors factorizes a single problem at a time. We illustrate how our performance analysis together with the profiling and tracing tools guided the development of batched factorizations to achieve up to 2-fold speedup and 3-fold better energy efficiency compared to our highly optimized batched CPU implementations based on the MKL library on a two-sockets, Intel Sandy Bridge server. Compared to a batched LU factorization featured in the NVIDIA's CUBLAS library for GPUs, we achieves up to 2.5-fold speedup on the K40 GPU.
This paper presents the design and implementation of several fundamental dense linear algebra (DLA) algorithms for multicore with Intel Xeon Phi coprocessors. In particular, we consider algorithms for solving linear systems. Further, we give an overview of the MAGMA MIC library, an open source, high performance library, that incorporates the developments presented here and, more broadly, provides the DLA functionality equivalent to that of the popular LAPACK library while targeting heterogeneous architectures that feature a mix of multicore CPUs and coprocessors. The LAPACK-compliance simplifies the use of the MAGMA MIC library in applications, while providing them with portably performant DLA. High performance is obtained through the use of the high-performance BLAS, hardware-specific tuning, and a hybridization methodology whereby we split the algorithm into computational tasks of various granularities. Execution of those tasks is properly scheduled over the heterogeneous hardware by minimizing data movements and mapping algorithmic requirements to the architectural strengths of the various heterogeneous hardware components. Our methodology and programming techniques are incorporated into the MAGMA MIC API, which abstracts the application developer from the specifics of the Xeon Phi architecture and is therefore applicable to algorithms beyond the scope of DLA.
Power and energy consumption are becoming an increasing concern in high performance computing. Compared to multi-core CPUs, GPUs have a much better performance per watt. In this paper we discuss efforts to redesign the most computation intensive parts of BLAST, an application that solves the equations for compressible hydrodynamics with high order finite elements, using GPUs BLAST, Dobrev. In order to exploit the hardware parallelism of GPUs and achieve high performance, we implemented custom linear algebra kernels. We intensively optimized our CUDA kernels by exploiting the memory hierarchy, which exceed the vendor's library routines substantially in performance. We proposed an auto tuning technique to adapt our CUDA kernels to the orders of the finite element method. Compared to a previous base implementation, our redesign and optimization lowered the energy consumption of the GPU in two aspects: 60% less time to solution and 10% less power required. Compared to the CPU-only solution, our GPU accelerated BLAST obtained a 2.5× overall speedup and 1.42× energy efficiency (green up) using 4th order (Q_4) finite elements, and a 1.9× speedup and 1.27× green up using 2nd order (Q2) finite elements.
Abstract not provided.
Abstract not provided.
The goal of the MAGMA project is to create a new generation of linear algebra libraries that achieve the fastest possible time to an accurate solution on hybrid Multicore+GPU-based systems, using all the processing power that future high-end systems can make available within given energy constraints. Our efforts at the University of Tennessee achieved the goals set in all of the five areas identified in the proposal: 1. Communication optimal algorithms; 2. Autotuning for GPU and hybrid processors; 3. Scheduling and memory management techniques for heterogeneity and scale; 4. Fault tolerance and robustness for large scale systems; 5. Building energy efficiency into software foundations. The University of Tennessee’s main contributions, as proposed, were the research and software development of new algorithms for hybrid multi/many-core CPUs and GPUs, as related to two-sided factorizations and complete eigenproblem solvers, hybrid BLAS, and energy efficiency for dense, as well as sparse, operations. Furthermore, as proposed, we investigated and experimented with various techniques targeting the five main areas outlined.
Abstract not provided.