Search NASA⌕ Search

SEARCH · Search NASA

Results for “Memory Hierarchy”

Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 109 records · Page 6

SU (2) lattice gauge theory simulations on Fermi GPUs

In this work we explore the performance of CUDA in quenched lattice SU (2) simulations. CUDA, NVIDIA Compute Unified Device Architecture, is a hardware and software architecture developed by NVIDIA for computing on the GPU. We present an analysis and performance comparison between the GPU and CPU in single and double precision. Analyses with multiple GPUs and two different architectures (G200 and Fermi architectures) are also presented. In order to obtain a high performance, the code must be optimized for the GPU architecture, i.e., an implementation that exploits the memory hierarchy of the CUDA programming model. We produce codes for the Monte Carlo generation of SU (2) lattice gauge configurations, for the mean plaquette, for the Polyakov Loop at finite T and for the Wilson loop. We also present results for the potential using many configurations (50,000) without smearing and almost 2000 configurations with APE smearing. With two Fermi GPUs we have achieved an excellent performance of 200x the speed over one CPU, in single precision, around 110 Gflops/s. We also find that, using the Fermi architecture, double precision computations for the static quark-antiquark potential are not much slower (less than 2x slower) than single precision computations.

72 PHYSICS OF ELEMENTARY PARTICLES AND FIELDS↗

Aspen-based performance and energy modeling frameworks

Now with the anticipation of exascale architectures, energy consumption is becoming one of the critical design parameters, especially in light of the energy budget of 20-30 Megawatts set by the U.S. Department of Energy. Understanding an application’s execution pattern and its energy footprint is critical to improving application operation on a diverse heterogeneous architecture. Applying application-specific performance optimization can consequently improve energy consumption. However, this approach is only applicable to current systems. As we enter a new era of exascale architectures that is projected to contain more complex memory hierarchies, increased levels of parallelism, heterogeneity in hardware, and complex programming models and techniques, energy and performance management is getting more cumbersome. We therefore propose techniques that predict the energy consumption beforehand or at runtime to enable proactive tuning. Such energy prediction approaches must be generic and adapt themselves at runtime to changing application and hardware configurations. Most existing energy estimation and prediction approaches are empirical in nature and thus tied to current systems. To overcome this limitation, we propose two energy estimation techniques: ACEE (Algorithmic and Categorical Energy Estimation), which uses a combination of analytical and empirical modeling techniques; and AEEM (Aspen’s Embedded Energy Estimation), a system-level analytical energy estimation technique. Both of these models incorporate the Aspen domain specific language for performance modeling. We present the methodologies of these two models and test their accuracy using five proxy applications. We also describe three use cases.

79 ASTRONOMY AND ASTROPHYSICS↗

FFT, FMM, and multigrid on the road to exascale: Performance challenges and opportunities

FFT, FMM, and multigrid methods are widely used fast and highly scalable solvers for elliptic PDEs. However, emerging large-scale computing systems are introducing challenges in comparison to current petascale computers. Recent efforts (Dongarra et al. 2011) have identified several constraints in the design of exascale software that include massive concurrency, resilience management, exploiting the high performance of heterogeneous systems, energy efficiency, and utilizing the deeper and more complex memory hierarchy expected at exascale. In this paper, we perform a model-based comparison of the FFT, FMM, and multigrid methods in the context of these projected constraints. Finally, we use performance models to offer predictions about the expected performance on upcoming exascale system configurations based on current technology trends.

97 MATHEMATICS AND COMPUTING↗

Basker: Parallel sparse LU factorization utilizing hierarchical parallelism and data layouts

Transient simulation in circuit simulation tools, such as SPICE and Xyce, depend on scalable and robust sparse LU factorizations for efficient numerical simulation of circuits and power grids. As the need for simulations of very large circuits grow, the prevalence of multicore architectures enable us to use shared memory parallel algorithms for such simulations. A parallel factorization is a critical component of such shared memory parallel simulations. We develop a parallel sparse factorization algorithm that can solve problems from circuit simulations efficiently, and map well to architectural features. This new factorization algorithm exposes hierarchical parallelism to accommodate irregular structure that arise in our target problems. It also uses a hierarchical two-dimensional data layout which reduces synchronization costs and maps to memory hierarchy found in multicore processors. We present an OpenMP based implementation of the parallel algorithm in a new multithreaded solver called Basker in the Trilinos framework. Here, we present performance evaluations of Basker on the Intel SandyBridge and Xeon Phi platforms using circuit and power grid matrices taken from the University of Florida sparse matrix collection and from Xyce circuit simulation. Basker achieves a geometric mean speedup of 5.91× on CPU (16 cores) and 7.4× on Xeon Phi (32 cores) relative to state-of-the-art solver KLU. Basker outperforms Intel MKL Pardiso solver (PMKL) by as much as 30× on CPU (16 cores) and 7.5× on Xeon Phi (32 cores) for low fill-in circuit matrices. Furthermore, Basker provides 5.4× speedup on a challenging matrix sequence taken from an actual Xyce simulation.

97 MATHEMATICS AND COMPUTING↗

Exploring HPCS Languages in Scientific Computing

As computers scale up dramatically to tens and hundreds of thousands of cores, develop deeper computational and memory hierarchies, and increased heterogeneity, developers of scientific software are increasingly challenged to express complex parallel simulations effectively and efficiently. In this paper, we explore the three languages developed under the DARPA High-Productivity Computing Systems (HPCS) program to help address these concerns: Chapel, Fortress, and X10. These languages provide a variety of features not found in currently popular HPC programming environments and make it easier to express powerful computational constructs, leading to new ways of thinking about parallel programming. Though the languages and their implementations are not yet mature enough for a comprehensive evaluation, we discuss some of the important features, and provide examples of how they can be used in scientific computing. We believe that these characteristics will be important to the future of high-performance scientific computing, whether the ultimate language of choice is one of the HPCS languages or something else.

parallel computing↗

A Code Generator for High-Performance Tensor Contractions on GPUs

Tensor contractions are higher dimensional generalizations of matrix-matrix multiplication. They form the compute-intensive core for many applications in computational science and data science. Although high-performance library implementations are widely available for dense matrix-matrix multiplication on GPUs, the same is not true for contraction of dense tensors. In this paper, we describe a high- performance GPU code generator for arbitrary tensor contractions. It exploits domain-specific properties about data reuse in tensor contractions to devise an effective code generation schema, coupled with an effective model-driven search to determine parameters for mapping of computation to threads and staging of data through the GPU memory hierarchy. Experimental evaluation using a set of tensor contraction benchmarks demonstrates performance improvement and/or significantly reduced code generation time over other state-of-the-art tensor contraction libraries and code generators.

Kim, Jinsung↗

Compiler generation and autotuning of communication-avoiding operators for geometric multigrid

This paper describes a compiler approach to introducing communication-avoiding optimizations in geometric multigrid (GMG), one of the most popular methods for solving partial differential equations. Communication-avoiding optimizations reduce vertical communication through the memory hierarchy and horizontal communication across processes or threads, usually at the expense of introducing redundant computation. We focus on applying these optimizations to the smooth operator, which successively reduces the error and accounts for the largest fraction of the GMG execution time. Our compiler technology applies both novel and known transformations to derive an implementation comparable to manually-tuned code. To make the approach portable, an underlying autotuning system explores the tradeoff between reduced communication and increased computation, as well as tradeoffs in threading schemes, to automatically identify the best implementation for a particular architecture and at each computation phase. Results show that we are able to quadruple the performance of the smooth operation on the finest grids while attaining performance within 94% of manually-tuned code. Overall we improve the overall multigrid solve time by 2.5× without sacrificing programer productivity.

97 MATHEMATICS AND COMPUTING↗

Parallel k-means Clustering of Geospatial Data Sets Using Manycore CPU Architectures

The increasing availability of high-resolution geospatiotemporal data sets from sources such as observatory networks, remote sensing platforms, and computational Earth system models has opened new possibilities for knowledge discovery and mining of weather, climate, ecological, and other geoscientific data sets fused from disparate sources. Many of the standard tools used on individual workstations are impractical for the analysis and synthesis of data sets of this size; however, new algorithmic approaches that can effectively utilize the complex memory hierarchies and the extremely high levels of parallelism available in state-of-the-art high-performance computing platforms can enable such analysis. Here, we describe pKluster, an open-source tool we have developed for accelerated k-means clustering of geospatial and geospatiotemporal data, and discuss algorithmic modifications and code optimizations we have made to enable it to effectively use parallel machines based on novel CPU architectures—such as the Intel “Knights Landing” Xeon Phi and Skylake Xeon processors—with many cores and hardware threads, and employing significant single instruction, multiple data (SIMD) parallelism. We outline some applications of the code in ecology and climate science contexts and present a detailed discussion of the performance of the code for one such application, LiDAR-derived vertical vegetation structure classification.

Mills, Richard T↗

A Step towards Energy Efficient Computing: Redesigning a Hydrodynamic Application on CPU-GPU, In: 2014 IEEE 28th International Parallel and Distributed Processing Symposium

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.

Computer Science↗

RAM as a Network Managed Resource

This paper describes an architecture for and several application examples of using a dynamically allocable RAM pool over the network as part of a deep memory hierarchy. We present four different use cases, including in situ analysis, machine learning, quantum chemistry simulations, and virtualization. In each use case, we modified or implemented software to use a network RAM pool, and then evaluated the performance. The cases of in situ analysis, machine learning, and virtualization demonstrated good performance. While the quantum chemistry experiments exhibited an average slowdown of 53% for all runs, including those where resources were clearly oversubscribed, the scalability results showed promise.

Allcock, William↗

PPT-GPU: Scalable GPU Performance Modeling

Performance modeling is a challenging problem due to the complexities of hardware architectures. In this paper, we present PPT-GPU, a scalable and accurate simulation framework that enables GPU code developers and architects to predict the performance of applications in a fast, and accurate manner on different GPU architectures. PPT-GPU is part of the open source project, Performance Prediction Toolkit (PPT) developed at the Los Alamos National Laboratory. We extend the old GPU model in PPT that predict the runtimes of computational physics codes to offer better prediction accuracy, for which, we add models for different memory hierarchies found in GPUs and latencies for different instructions. To further show the utility of PPT-GPU, we compare our model against real GPU device(s) and the widely used cycle-accurate simulator, GPGPU-Sim using different workloads from RODINIA and Parboil benchmarks. The results indicate that the predicted performance of PPT-GPU is within a 10% error compared to the real device(s). In addition, PPT-GPU is highly scalable, where it is up to 450x faster than GPGPU-Sim with more accurate results.

97 MATHEMATICS AND COMPUTING↗

The Persistent Challenge of Data Locality in the Post-Exascale Era

The era of exascale computing, exemplified by systems like Frontier achieving exaflop-level performance, marks a milestone. However, the quest for sheer compute power leads to strong imbalance in system design. Hence, scaling advancements in memory, network bandwidth, and storage are also necessary and pose challenges, with a crucial need to address data locality issues. This article underscores the fundamental importance of data locality as a key abstraction for optimizing application performance. Despite notable software solutions, the growing complexity of parallelism and memory hierarchy demands performance-portable data locality solutions across diverse computing platforms. Additionally, the article revisits data locality aspects, covering hardware considerations, application perspectives, software stack abstractions, and tool support. It concludes with insights into data locality challenges and opportunities, emphasizing the ongoing significance of collaborative research for progress in this critical issue.

Unat, Didem [Koc University, Istanbul (Turkey)] (O↗

Automatic Code Generation for High-Performance Graph Algorithms

Graph problems are common across fields of scientific computing and social sciences. However, despite their importance, implementing graph algorithms effectively on modern computing systems is a challenging task that requires significant programming effort and generally results in customized implementations. Current computing and memory hierarchies are not architected for irregular computations resulting in challenges for graph algorithms to achieve high performance on those architectures. In this paper, we present GraphX, a novel compiler framework and DSL designed to simplify the development of efficient graph algorithms and achieve high performance on modern computing systems. GraphX consists of a DSL for efficient implementation of graph algorithms, various optimizations, such as support for sparse linear algebra and workspace transformations, optimized graph primitives, including semiring and masking, and a high-performance code generation engine. Using GraphX, users can implement graph algorithms using a semantically-rich language with graph-oriented operators. GraphX uses these semantics to automatically generate efficient code for target architectures, increasing performance and portability across architectures. The composable nature of GraphX makes it possible to extend the set of optimizations and architectures without modifying the source code. We demonstrate GraphX outperforms state-of-the-art graph libraries, such as LAGraph, up to $3.7 speedup in semiring operations, $2.19 speedup in an important sparse computational kernel, and $9.05 speedup in graph processing algorithms.

compiler, graph algorithms, semiring, masking, wor↗

Trends in data locality abstractions for HPC systems

The cost of data movement has always been an important concern in high performance computing (HPC) systems. It has now become the dominant factor in terms of both energy consumption and performance. Support for expression of data locality has been explored in the past, but those efforts have had only modest success in being adopted in HPC applications for various reasons. However, with the increasing complexity of the memory hierarchy and higher parallelism in emerging HPC systems, locality management has acquired a new urgency. Developers can no longer limit themselves to low-level solutions and ignore the potential for productivity and performance portability obtained by using locality abstractions. Fortunately, the trend emerging in recent literature on the topic alleviates many of the concerns that got in the way of their adoption by application developers. Data locality abstractions are available in the forms of libraries, data structures, languages and runtime systems; a common theme is increasing productivity without sacrificing performance. Furthermore, this paper examines these trends and identifies commonalities that can combine various locality concepts to develop a comprehensive approach to expressing and managing data locality on future large-scale high-performance computing systems.

97 MATHEMATICS AND COMPUTING↗

Communication-avoiding symmetric-indefinite factorization

We describe and analyze a novel symmetric triangular factorization algorithm. The algorithm is essentially a block version of Aasen's triangular tridiagonalization. It factors a dense symmetric matrix A as the product A=PLTL T P T where P is a permutation matrix, L is lower triangular, and T is block tridiagonal and banded. The algorithm is the first symmetric-indefinite communication-avoiding factorization: it performs an asymptotically optimal amount of communication in a two-level memory hierarchy for almost any cache-line size. Adaptations of the algorithm to parallel computers are likely to be communication efficient as well; one such adaptation has been recently published. As a result, the current paper describes the algorithm, proves that it is numerically stable, and proves that it is communication optimal.

97 MATHEMATICS AND COMPUTING↗

Parameterized Micro-benchmarking: An Auto-tuning Approach for Complex Applications

Auto-tuning has emerged as an important practical method for creating highly optimized implementations of key computational kernels and applications. However, the growing complexity of architectures and applications is creating new challenges for auto-tuning. Complex applications can involve a prohibitively large search space that precludes empirical auto-tuning. Similarly, architectures are becoming increasingly complicated, making it hard to model performance. In this paper, we focus on the challenge to auto-tuning presented by applications with a large number of kernels and kernel instantiations. While these kernels may share a somewhat similar pattern, they differ considerably in problem sizes and the exact computation performed. We propose and evaluate a new approach to auto-tuning which we refer to as parameterized micro-benchmarking. It is an alternative to the two existing classes of approaches to auto-tuning: analytical model-based and empirical search-based. Particularly, we argue that the former may not be able to capture all the architectural features that impact performance, whereas the latter might be too expensive for an application that has several different kernels. In our approach, different expressions in the application, different possible implementations of each expression, and the key architectural features, are used to derive a simple micro-benchmark and a small parameter space. This allows us to learn the most significant features of the architecture that can impact the choice of implementation for each kernel. We have evaluated our approach in the context of GPU implementations of tensor contraction expressions encountered in excited state calculations in quantum chemistry. We have focused on two aspects of GPUs that affect tensor contraction execution: memory access patterns and kernel consolidation. Using our parameterized micro-benchmarking approach, we obtain a speedup of up to 2 over the version that used default optimizations, but no auto-tuning. We demonstrate that observations made from microbenchmarks match the behavior seen from real expressions. In the process, we make important observations about the memory hierarchy of two of the most recent NVIDIA GPUs, which can be used in other optimization frameworks as well.

auto-tuning↗

B-Trees and Cache-Oblivious B-Trees with Different-Sized Atomic Keys

Most B-tree articles assume that all $\textit{N}$ keys have the same size $\textit{K}$, that $\textit{f = B/K}$ keys fit in a disk block, and therefore that the search cost is $O(\text{log}_{f+1}N) $ block transfers. When keys have variable size, B-tree operations have no nontrivial performance guarantees, however. This article provides B-tree-like performance guarantees on dictionaries that contain keys of different sizes in a model in which keys must be stored and compared as opaque objects. The resulting atomic-key dictionaries exhibit performance bounds in terms of the average key size and match the bounds when all keys are the same size. Atomic-key dictionaries can be built with minimal modification to the B-tree structure, simply by choosing the pivot keys properly. Further, this article describes both static and dynamic atomic-key dictionaries. In the static case, if there are $\textit{N}$ keys with average size $\textit{K}$, the search cost is $O(\lceil K/B\rceil \text{log}_{1 + \lceil K/B\rceil} N)$ expected transfers. It is not possible to transform these expected bounds into worst-case bounds. The cost to build the tree is $\textit{O(NK)}$ operations and $\textit{O(NK/B)}$ transfers if all keys are presented in sorted order. If not, the cost is the sorting cost. For the dynamic dictionaries, the amortized cost to insert a key $\textit{κ}$ of arbitrary length at an arbitrary rank is dominated by the cost to search for $\textit{κ}$. Specifically, the amortized cost to insert a key $\textit{κ}$ of arbitrary length and random rank is $O(\lceil K/B\rceil log_{1+ \lceil B/K\rceil } N + |κ|/B)$ transfers. A dynamic-programming algorithm is shown for constructing a search tree with minimal expected cost. This article also gives a cache-oblivious static atomic-key B-tree, which achieves the same asymptotic performance as the static B-tree dictionary, mentioned previously. A cache-oblivious data structure or algorithm is not parameterized by the block size $\textit{B}$ or memory size $\textit{M}$ in the memory hierarchy; rather, it is universal, working simultaneously for all possible values of $\textit{B}$ or $\textit{M}$. On a machine with block size $\textit{B}$, if there are $\textit{N}$ keys with average size $\textit{K}$, search operations costs $O(\lceil K/B \rceil log_{1+ \lceil B/K \rceil} N)$ block transfers in expectation. This cache-oblivious layout can be built in $\textit{O(N}\text{log}\textit{(NK))}$ processor operations.

97 MATHEMATICS AND COMPUTING↗

Data and Thread Placement in NUMA Architectures: A Statistical Learning Approach

Nowadays, NUMA architectures are common in compute-intensive systems. Achieving high performance for multi-threaded application requires both a careful placement of threads on computing units and a thorough allocation of data in memory. Finding such a placement is a hard problem to solve, because performance depends on complex interactions in several layers of the memory hierarchy. In this paper we propose a black-box approach to decide if an application execution time can be impacted by the placement of its threads and data, and in such a case, to choose the best placement strategy to adopt. We show that it is possible to reach near-optimal placement policy selection. Furthermore, solutions work across several recent processor architectures and decisions can be taken with a single run of low overhead profiling.

data↗