Engineering Papers⌕ Search

Engineering topics

Shalf, John

Publications and source records attributed to Shalf, John.

At least 73 records · Page 4

Auto-Tuning Memory-Intensive Kernels for Multicore

In this, chapter, we discuss the optimization of three memory-intensive computational kernels — sparse matrix-vector multiplication, the Laplacian differential operator applied to structured grids, and the collision() operator with the lattice Boltzmann magnetohydrodynamics (LBMHD) application. They are all implemented using a single-process, (POSIX) threaded, SPMD model. Unlike their computationally-intense dense linear algebra cousins, performance is ultimately limited by DRAM bandwidth and the volume of data that must be transfered. To provide performance portability across current and future multicore architectures, we utilize automatic performance tuning, or auto-tuning.

97 MATHEMATICS AND COMPUTING↗

Tuning HDF5 for Lustre File Systems

HDF5 is a cross-platform parallel I/O library that is used by a wide variety of HPC applications for the flexibility of its hierarchical object-database representation of scientific data. We describe our recent work to optimize the performance of the HDF5 and MPI-IO libraries for the Lustre parallel file system. We selected three different HPC applications to represent the diverse range of I/O requirements, and measured their performance on three different systems to demonstrate the robustness of our optimizations across different file system configurations and to validate our optimization strategy. We demonstrate that the combined optimizations improve HDF5 parallel I/O performance by up to 33 times in some cases running close to the achievable peak performance of the underlying file system and demonstrate scalable performance up to 40,960-way concurrency.

97 MATHEMATICS AND COMPUTING↗

A Testbed of Parallel Kernels for Computer Science Research

For several decades, computer scientists have sought guidance on how to evolve architectures, languages, and programming models for optimal performance, efficiency, and productivity. Unfortunately, this guidance is most often taken from the existing software/hardware ecosystem. Architects attempt to provide micro-architectural solutions to improve performance on fixed binaries. Researchers tweak compilers to improve code generation for existing architectures and implementations, and they may invent new programming models for fixed processor and memory architectures and computational algorithms. In today's rapidly evolving world of on-chip parallelism, these isolated and iterative improvements to performance may miss superior solutions in the same way gradient descent optimization techniques may get stuck in local minima. In an initial study, we have developed an alternate approach that, rather than starting with an existing hardware/software solution laced with hidden assumptions, defines the computational problems of interest and invites architects, researchers and programmers to implement novel hardware/ software co-designed solutions. Our work builds on the previous ideas of computational dwarfs, motifs, and parallel patterns by selecting a representative set of essential problems for which we provide: An algorithmic description; scalable problem definition; illustrative reference implementations; verification schemes. For simplicity, we focus initially on the computational problems of interest to the scientific computing community but proclaim the methodology (and perhaps a subset of the problems) as applicable to other communities. We intend to broaden the coverage of this problem space through stronger community involvement. Previous work has established a broad categorization of numerical methods of interest to the scientific computing, in the spirit of the NAS Benchmarks, which pioneered the basic idea of a 'pencil and paper benchmark' in the 1990s. The initial result of the more modern study was the seven dwarfs, which was subsequently extended to 13 motifs. These motifs have already been useful in defining classes of applications for architecture-software studies. However, these broad-brush problem statements often miss the nuance seen in individual kernels. For example, the computational requirements of particle methods vary greatly between the naive (but more accurate) direct calculations and the particle-mesh and particle-tree codes. Thus we commenced our study with an enumeration of problems, but then proceeded by providing not only reference implementations for each problem, but more importantly a mathematical definition that allows one to escape iterative approaches to software/hardware optimization. To ensure long term value, we have augmented each of our reference implementations with both a scalable problem generator and a verification scheme. In a paper we have prepared that documents our efforts, we describe in detail this process of problem definition, scalable input creation, verification, and implementation of reference codes for the scientific computing domain. Table 1 enumerates and describes the level of support we've developed for each kernel. We group these important kernels using the Berkeley dwarfs/motifs taxonomy using a red box in the appropriate column. As kernels become progressively complex, they build upon other, simpler computational methods. We note this dependency via orange boxes. After enumeration of the important numerical problems, we created a domain-appropriate high-level definition of each problem. To ensure future endeavors are not tainted by existing implementations, we specified the problem definition to be independent of both computer architecture and existing programming languages, models, and data types. Then, to provide context as to how such kernels productively map to existing architectures, languages and programming models, we produced reference implementations for most of the kernel (see table). These sample codes should be viewed as 'hints,' designed to show how other designers have mapped a problem's operands and operators to existing hardware and software. Since we wanted such implementations to be illustrative, we tried to ensure they were the most straightforward implementation in the easiest to understand languages using familiar architectures. To that end, most of the linear algebra-oriented computations are written in MATLAB using array indexing to process matrices, rather than one-line library calls to compute the same kernel. This ensures that the kernel's computation is explicit and readable in the implementation and not hidden behind a library. For other problems, such as the Barnes-Hut n-Body solver, the implementations were written in pure C, without any supporting library computations. Along this line, we have also created a scalable problem generator to accompany each computation.

97 MATHEMATICS AND COMPUTING↗

Auto-Tuning the 27-point Stencil for Multicore

This study focuses on the key numerical technique of stencil computations, used in many different scientific disciplines, and illustrates how auto-tuning can be used to produce very efficient implementations across a diverse set of current multicore architectures.

97 MATHEMATICS AND COMPUTING↗

A Generalized Framework for Auto-tuning Stencil Computations

This work introduces a generalized framework for automatically tuning stencil computations to achieve superior performance on a broad range of multicore architectures. Stencil (nearest-neighbor) based kernels constitute the core of many important scientific applications involving block-structured grids. Auto-tuning systems search over optimization strategies to find the combination of tunable parameters that maximizes computational efficiency for a given algorithmic kernel. Although the auto-tuning strategy has been successfully applied to libraries, generalized stencil kernels are not amenable to packaging as libraries. Studied kernels in this work include both memory-bound kernels as well as a computation-bound bilateral filtering kernel. We introduce a generalized stencil auto-tuning framework that takes a straightforward Fortran expression of a stencil kernel and automatically generates tuned implementations of the kernel in C or Fortran to achieve performance portability across diverse computer architectures.

97 MATHEMATICS AND COMPUTING↗

Resource-Efficient, Hierarchical Auto-Tuning of a Hybrid Lattice Boltzmann Computation on the Cray XT4

We apply auto-tuning to a hybrid MPI-pthreads lattice Boltzmann computation running on the Cray XT4 at National Energy Research Scientific Computing Center (NERSC). Previous work showed that multicore-specific auto-tuning can improve the performance of lattice Boltzmann magnetohydrodynamics (LBMHD) by a factor of 4x when running on dual- and quad-core Opteron dual-socket SMPs. We extend these studies to the distributed memory arena via a hybrid MPI/pthreads implementation. In addition to conventional auto-tuning at the local SMP node, we tune at the message-passing level to determine the optimal aspect ratio as well as the correct balance between MPI tasks and threads per MPI task. Our study presents a detailed performance analysis when moving along an isocurve of constant hardware usage: fixed total memory, total cores, and total nodes. Overall, our work points to approaches for improving intra- and inter-node efficiency on large-scale multicore systems for demanding scientific applications.

97 MATHEMATICS AND COMPUTING↗

A Generalized Framework for Auto-tuning Stencil Computations

This work introduces a generalized framework for automatically tuning stencil computations to achieve superior performance on a broad range of multicore architectures. Stencil (nearest-neighbor) based kernels constitute the core of many important scientific applications involving block-structured grids. Auto-tuning systems search over optimization strategies to find the combination of tunable parameters that maximizes computational efficiency for a given algorithmic kernel. Although the auto-tuning strategy has been successfully applied to libraries, generalized stencil kernels are not amenable to packaging as libraries. Studied kernels in this work include both memory-bound kernels as well as a computation-bound bilateral filtering kernel. We introduce a generalized stencil auto-tuning framework that takes a straightforward Fortran expression of a stencil kernel and automatically generates tuned implementations of the kernel in C or Fortran to achieve performance portability across diverse computer architectures.

97 MATHEMATICS AND COMPUTING↗

Optimization of a Lattice Boltzmann Computation on State-of-the-Art Multicore Platforms

We present an auto-tuning approach to optimize application performance on emerging multicore architectures. The methodology extends the idea of search-based performance optimizations, popular in linear algebra and FFT libraries, to application-specific computational kernels. Our work applies this strategy to a lattice Boltzmann application (LBMHD) that historically has made poor use of scalar microprocessors due to its complex data structures and memory access patterns. We explore one of the broadest sets of multicore architectures in the HPC literature, including the Intel Xeon E5345 (Clovertown), AMD Opteron 2214 (Santa Rosa), AMD Opteron 2356 (Barcelona), Sun T5140 T2+ (Victoria Falls), as well as a QS20 IBM Cell Blade. Rather than hand-tuning LBMHD for each system, we develop a code generator that allows us to identify a highly optimized version for each platform, while amortizing the human programming effort. Results show that our auto-tuned LBMHD application achieves up to a 15x improvement compared with the original code at a given concurrency. Additionally, we present detailed analysis of each optimization, which reveal surprising hardware bottlenecks and software challenges for future multicore systems and applications.

97 MATHEMATICS AND COMPUTING↗

Analysis of Photonic Networks for a Chip Multiprocessor Using Scientific Applications

As multiprocessors scale to unprecedented numbers of cores in order to sustain performance growth, it is vital that these gains are not nullified by high energy consumption from inter-core communication. With recent advances in 3D Integration CMOS technology, the possibility for realizing hybrid photonic-electronic networks-on-chip warrants investigating real application traces on functionally comparable photonic and electronic network designs. We present a comparative analysis using both synthetic benchmarks as well as real applications, run through detailed cycle accurate models implemented under the OMNeT++ discrete event simulation environment. Results show that when utilizing standard process-to-processor mapping methods, this hybrid network can achieve 75X improvement in energy efficiency for synthetic benchmarks and up to 37X improvement for real scientific applications, defined as network performance per energy spent, over an electronic mesh for large messages across a variety of communication patterns.

97 MATHEMATICS AND COMPUTING↗

Improving Memory Subsystem Performance Using ViVA: Virtual Vector Architecture

The disparity between microprocessor clock frequencies and memory latency is a primary reason why many demanding applications run well below peak achievable performance. Software controlled scratchpad memories, such as the Cell local store, attempt to ameliorate this discrepancy by enabling precise control over memory movement; however, scratchpad technology confronts the programmer and compiler with an unfamiliar and difficult programming model. In this work, we present the Virtual Vector Architecture (ViVA), which combines the memory semantics of vector computers with a software-controlled scratchpad memory in order to provide a more effective and practical approach to latency hiding. ViVA requires minimal changes to the core design and could thus be easily integrated with conventional processor cores. To validate our approach, we implemented ViVA on the Mambo cycle-accurate full system simulator, which was carefully calibrated to match the performance on our underlying PowerPC Apple G5 architecture. Results show that ViVA is able to deliver significant performance benefits over scalar techniques for a variety of memory access patterns as well as two important memory-bound compact kernels, corner turn and sparse matrix-vector multiplication -- achieving 2x-13x improvement compared the scalar version. Overall, our preliminary ViVA exploration points to a promising approach for improving application performance on leading microprocessors with minimal design and complexity costs, in a power efficient manner.

97 MATHEMATICS AND COMPUTING↗

A design methodology for domain-optimized power-efficient supercomputing

As power has become the pre-eminent design constraint for future HPC systems, computational efficiency is being emphasized over simply peak performance. Recently, static benchmark codes have been used to find a power efficient architecture. Unfortunately, because compilers generate sub-optimal code, benchmark performance can be a poor indicator of the performance potential of architecture design points. Therefore, we present hardware/software cotuning as a novel approach for system design, in which traditional architecture space exploration is tightly coupled with software auto-tuning for delivering substantial improvements in area and power efficiency. We demonstrate the proposed methodology by exploring the parameter space of a Tensilica-based multi-processor running three of the most heavily used kernels in scientific computing, each with widely varying micro-architectural requirements: sparse matrix vector multiplication, stencil-based computations, and general matrix-matrix multiplication. Results demonstrate that co-tuning significantly improves hardware area and energy efficiency - a key driver for next generation of HPC system design.

97 MATHEMATICS AND COMPUTING↗

Memory-efficient optimization of Gyrokinetic particle-to-grid interpolation for multicore processors

We present multicore parallelization strategies for the particle-to-grid interpolation step in the Gyrokinetic Toroidal Code (GTC), a 3D particle-in-cell (PIC) application to study turbulent transport in magnetic-confinement fusion devices. Particle-grid interpolation is a known performance bottleneck in several PIC applications. In GTC, this step involves particles depositing charges to a 3D toroidal mesh, and multiple particles may contribute to the charge at a grid point. We design new parallel algorithms for the GTC charge deposition kernel, and analyze their performance on three leading multicore platforms. We implement thirteen different variants for this kernel and identify the best-performing ones given typical PIC parameters such as the grid size, number of particles per cell, and the GTC-specific particle Larmor radius variation. We find that our best strategies can be 2x faster than the reference optimized MPI implementation, and our analysis provides insight into desirable architectural features for high-performance PIC simulation codes.

97 MATHEMATICS AND COMPUTING↗

Optimization of Sparse Matrix-Vector Multiplication on Emerging Multicore Platforms

We are witnessing a dramatic change in computer architecture due to the multicore paradigm shift, as every electronic device from cell phones to supercomputers confronts parallelism of unprecedented scale. To fully unleash the potential of these systems, the HPC community must develop multicore specific-optimization methodologies for important scientific computations. In this work, we examine sparse matrix-vector multiply (SpMV) - one of the most heavily used kernels in scientific computing - across a broad spectrum of multicore designs. Our experimental platform includes the homogeneous AMD quad-core, AMD dual-core, and Intel quad-core designs, the heterogeneous STI Cell, as well as one of the first scientific studies of the highly multithreaded Sun Victoria Falls (a Niagara2 SMP). We present several optimization strategies especially effective for the multicore environment, and demonstrate significant performance improvements compared to existing state-of-the-art serial and parallel SpMV implementations. Additionally, we present key insights into the architectural trade-offs of leading multicore design strategies, in the context of demanding memory-bound numerical algorithms.

97 MATHEMATICS AND COMPUTING↗

NERSC-6 Workload Analysis and Benchmark Selection Process

This report describes efforts carried out during early 2008 to determine some of the science drivers for the"NERSC-6" next-generation high-performance computing system acquisition. Although the starting point was existing Greenbooks from DOE and the NERSC User Group, the main contribution of this work is an analysis of the current NERSC computational workload combined with requirements information elicited from key users and other scientists about expected needs in the 2009-2011 timeframe. The NERSC workload is described in terms of science areas, computer codes supporting research within those areas, and description of key algorithms that comprise the codes. This work was carried out in large part to help select a small set of benchmark programs that accurately capture the science and algorithmic characteristics of the workload. The report concludes with a description of the codes selected and some preliminary performance data for them on several important systems.

97 MATHEMATICS AND COMPUTING↗

HPC Global File System Performance Analysis Using A Scientific-Application Derived Benchmark

With the exponential growth of high-fidelity sensor and simulated data, the scientific community is increasingly reliant on ultrascale HPC resources to handle its data analysis requirements. However, to use such extreme computing power effectively, the I/O components must be designed in a balanced fashion, as any architectural bottleneck will quickly render the platform intolerably inefficient. To understand I/O performance of data-intensive applications in realistic computational settings, we develop a lightweight, portable benchmark called MADbench2, which is derived directly from a large-scale Cosmic Microwave Background (CMB) data analysis package. Our study represents one of the most comprehensive I/O analyses of modern parallel file systems, examining a broad range of system architectures and configurations, including Lustre on the Cray XT3, XT4, and Intel Itanium2 clusters; GPFS on IBM Power5 and AMD Opteron platforms; a BlueGene/P installation using GPFS and PVFS2 file systems; and CXFS on the SGI Altix\-3700. We present extensive synchronous I/O performance data comparing a number of key parameters including concurrency, POSIX- versus MPI-IO, and unique-versus shared-file accesses, using both the default environment as well as highly-tuned I/O parameters. Finally, we explore the potential of asynchronous I/O and show that only the two of the nine evaluated systems benefited from MPI-2's asynchronous MPI-IO. On those systems, experimental results indicate that the computational intensity required to hide I/O effectively is already close to the practical limit of BLAS3 calculations. Overall, our study quantifies vast differences in performance and functionality of parallel file systems across state-of-the-art platforms -- showing I/O rates that vary up to 75x on the examined architectures -- while providing system designers and computational scientists a lightweight tool for conducting further analysis.

97 MATHEMATICS AND COMPUTING↗

Stencil Computation Optimization and Auto-tuning on State-of-the-Art Multicore Architectures

Understanding the most efficient design and utilization of emerging multicore systems is one of the most challenging questions faced by the mainstream and scientific computing industries in several decades. Our work explores multicore stencil (nearest-neighbor) computations -- a class of algorithms at the heart of many structured grid codes, including PDE solvers. We develop a number of effective optimization strategies, and build an auto-tuning environment that searches over our optimizations and their parameters to minimize runtime, while maximizing performance portability. To evaluate the effectiveness of these strategies we explore the broadest set of multicore architectures in the current HPC literature, including the Intel Clovertown, AMD Barcelona, Sun Victoria Falls, IBM QS22 PowerXCell 8i, and NVIDIA GTX280. Overall, our auto-tuning optimization methodology results in the fastest multicore stencil performance to date. Finally, we present several key insights into the architectural trade-offs of emerging multicore designs and their implications on scientific algorithm development.

97 MATHEMATICS AND COMPUTING↗

The roofline model: A pedagogical tool for program analysis and optimization

This article consists of a collection of slides from the authors' conference presentation. The Roofline model is a visually intuitive figure for kernel analysis and optimization. We believe undergraduates will find it useful in assessing performance and scalability limitations. It is easily extended to other architectural paradigms. It is easily extendable to other metrics: performance (sort, graphics, crypto..) bandwidth (L2, PCIe, ..). Furthermore, a performance counters could be used to generate a runtime-specific roofline that would greatly aide the optimization.

97 MATHEMATICS AND COMPUTING↗

PERI - Auto-tuning Memory Intensive Kernels for Multicore

We present an auto-tuning approach to optimize application performance on emerging multicore architectures. The methodology extends the idea of search-based performance optimizations, popular in linear algebra and FFT libraries, to application-specific computational kernels. Our work applies this strategy to Sparse Matrix Vector Multiplication (SpMV), the explicit heat equation PDE on a regular grid (Stencil), and a lattice Boltzmann application (LBMHD). We explore one of the broadest sets of multicore architectures in the HPC literature, including the Intel Xeon Clovertown, AMD Opteron Barcelona, Sun Victoria Falls, and the Sony-Toshiba-IBM (STI) Cell. Rather than hand-tuning each kernel for each system, we develop a code generator for each kernel that allows us to identify a highly optimized version for each platform, while amortizing the human programming effort. Results show that our auto-tuned kernel applications often achieve a better than 4X improvement compared with the original code. Additionally, we analyze a Roofline performance model for each platform to reveal hardware bottlenecks and software challenges for future multicore systems and applications.

97 MATHEMATICS AND COMPUTING↗