Design of a vector processor

This paper discusses the inherent parallelism limits on several applications for vector computers, the parallel capabilities of several architectures and two ways (traditional instruction control flow and data control flow) by which the capabilities can be used. Then a scheme for a pipelined vector processor of multi-processing units is presented. The basic system structure and its function on highly sparse vector processing are described. A vector cache system and a distributed main memory are also considered, which are intended to sustain extremely high access rates for the processor. A microprocessor based vector processor is constructed, which can simulate the high performance version of the processor.     

Region-based parallelization of irregular reductions on explicitly managed memory hierarchies

Multicore architectures are evolving with the promise of extreme performance for the classes of applications that require high performance and large bandwidth of memory. Irregular reduction is one of important computation patterns for many complex scientific applications, and it typically requires high performance and large bandwidth of memory. In this article, we propose region-based parallelization techniques for irregular reductions on multicore architectures with explicitly managed memory hierarchies. Managing memory hierarchy in software requires a lot of programming efforts and tends to be error-prone. The difficulties are even worse for applications with irregular data access patterns. To relieve the burden of memory management from programmers, we develop abstractions, particularly targeted to irregular reduction, for structuring parallel tasks, mapping the parallel tasks to processing units and scheduling data transfers between the memory hierarchies. Our framework employs iteration reordering based on regions of data along with dynamic scheduling of parallel tasks. We experimentally evaluate the effectiveness of our techniques for irregular reduction kernels on the Cell processor embedded in a Sony PlayStation3. Experimental results show the speedups of 8 to 14 on the six available SPEs.     

Compilation techniques for parallel systems

Over the past two decades tremendous progress has been made in both the design of parallel architectures and the compilers needed for exploiting parallelism on such architectures. In this paper we summarize the advances in compilation techniques for uncovering and effectively exploiting parallelism at various levels of granularity. We begin by describing the program analysis techniques through which parallelism is detected and expressed in form of a program representation. Next compilation techniques for scheduling instruction level parallelism (ILP) are discussed along with the relationship between the nature of compiler support and type of processor architecture. Compilation techniques for exploiting loop and task level parallelism on shared-memory multiprocessors (SMPs) are summarized. Locality optimizations that must be used in conjunction with parallelization techniques for achieving high performance on machines with complex memory hierarchies are also discussed. Finally we provide an overview of compilation techniques for distributed memory machines that must perform partitioning of both code and data for parallel execution. Communication optimization and code generation issues that are unique to such compilers are also briefly discussed.     

Power/Performance/Thermal Design-Space Exploration for Multicore Architectures

Multicore architectures have been ruling the recent microprocessor design trend. This is due to different reasons: better performance, thread-level parallelism bounds in modern applications, ILP diminishing returns, better thermal/power scaling (many small cores dissipate less than a large and complex one), and the ease and reuse of design. This paper presents a thorough evaluation of multicore architectures. The architecture that we target is composed of a configurable number of cores, a memory hierarchy consisting of private L1, shared/private L2, and a shared bus interconnect. We consider a benchmark set composed of several parallel shared memory applications. We explore the design space related to the number of cores, L2 cache size, and processor complexity, showing the behavior of the different configurations/applications with respect to performance, energy consumption, and temperature. Design trade-offs are analyzed, stressing the interdependency of the metrics and design factors. In particular, we evaluate several chip floorplans. Their power/thermal characteristics are analyzed, showing the importance of considering thermal effects at the architectural level to achieve the best design choice.     

Performance of symbolic applications on a parallel architecture

The results of a study of a family of parallel symbolic architectures executing several parallel applications are presented. The class of architectures being simulated is characterized by a shared memory structure, by a hierarchical interconnect, and by clustered processors. Speedup measurements were obtained from six different application kernels. Measurements were also performed to assess the degradation of speedup as a function of the interconnection delays, and to study the effect of different scheduling algorithms. The results presented support the claim that the proposed architecture would be a powerful parallel symbolic computation system. The paper discusses processor starvation, fine grain parallelism, unever loads, foreign reference, schedule and indeterminate computation with respect to the applications chosen.This work was completed within the Advanced Computer Architecture Program, Micro-electronics and Technology Computer Corporation, Austin, Texas.     

Controlling NUMA effects in embedded manycore applications with lightweight nested parallelism support

Embedded manycore architectures are often organized as fabrics of tightly-coupled shared memory clusters. A hierarchical interconnection system is used with a crossbar-like medium inside each cluster and a network-on-chip (NoC) at the global level which make memory operations nonuniform (NUMA). Due to NUMA, regular applications typically employed in the embedded domain (e.g., image processing, computer vision, etc.) ultimately behave as irregular workloads if a flat memory system is assumed at the program level. Nested parallelism represents a powerful programming abstraction for these architectures, provided that (i) streamlined middleware support is available, whose overhead does not dominate the run-time of fine-grained applications; (ii) a mechanism to control thread binding at the cluster-level is supported. We present a lightweight runtime layer for nested parallelism on cluster-based embedded manycores, integrating our primitives in the OpenMP runtime system, and implementing a new directive to control NUMA-aware nested parallelism mapping. We explore on a set of real application use cases how NUMA makes regular parallel workloads behave as irregular, and how our approach allows to control such effects and achieve up to 28 × speedup versus flat parallelism.     

A programming methodology for dual-tier multicomputers

Hierarchically organized ensembles of shared memory multiprocessors possess a richer and more complex model of locality than previous generation multicomputers with single processor nodes. These dual-tier computers introduce many new factors into the programmer's performance model. We present a methodology for implementing block-structured numerical applications on dual-tier computers and a run-time infrastructure, called KeLP2, that implements the methodology. KeLP2 supports two levels of locality and parallelism via hierarchical SPMD control flow, run-time geometric meta-data, and asynchronous collective communication. KeLP applications can effectively overlap communication with computation under conditions where nonblocking point-to-point message passing fails to do so. KeLP's abstractions hide considerable detail without sacrificing performance and dual-tier applications written in KeLP consistently outperform equivalent single-tier implementations written in MPI. We describe the KeLP2 model and show how it facilitates the implementation of five block-structured applications specially formulated to hide communication latency on dual-tiered architectures. We support our arguments with empirical data from applications running on various single- and dual-tier multicomputers. KeLP2 supports a migration path from single-tier to dual-tier platforms and we illustrate this capability with a detailed programming example     

X处理器存储层次研究

随着计算机应用领域不断拓展,流媒体应用及科学计算正成为微处理器的一种重要负载.流媒体应用的特征是大量的数据并行、少量的数据重用以及每次访存带来的大量计算.因为带宽的限制,传统的微处理器结构很难满足这些特点.X处理器是一款流处理器,针对流应用特点,X处理器采用了新型的三级流式存储层次:局部寄存器文件、流寄存器文件和片外存储器,有效解决了带宽问题.本文在模拟平台采用了两种方法(RS码和测试程序)测试,验证了流存储层次解决带宽瓶颈的有效性,也证明了设计的正确性.     

COMBINING BACKGROUND MEMORY MANAGEMENT AND REGULAR ARRAY CO-PARTITIONING,ILLUSTRATED ON A FULL MOTION ESTIMATION KERNEL

Abstract

In this paper an approach is presented to combine the design of background memory architectures and processor arrays for data dominated real-time applications. The formalized data transfer and storage exploration (DTSE) approach of IMEC involves a methodology for the design of a low-power small-size background memory organizations, meeting real-time constraints. The systematic space-time transformation and the subsequent co-partitioning approach of the Dresden University of Technology, allow the design of realistic processor arrays adapted to a given memory architecture. However, neither methodology can derive on its own the complete solution of a fully optimized memory organization, combining background and foreground memory. Extensions to enable this important problem will be presented here. First, both complementary methodologies will be summarized. Next, the main emphasis in this paper will be on the approach to design the processor array within the context of an already optimized and hence given memory architecture. The feasibility of the proposed combination is demonstrated on a representative test-vehicle for an important class of applications, namely a full motion estimation kernel in MPEG.     

科学计算程序在FT64流处理器上的实现、优化和评测

流体系结构是一种适应VLSI工艺发展的新型体系结构,它是否对科学计算程序有效是一个广泛关注的问题。本文选取NASA并行测试程序集中的一个数据密集型程序MG,研究了 它在一个64位的面向科学计算设计的流处理器FT64上的实现和优化问题。在FT64上的实测表明,经过面向片上存储层次的优化,FT64能够达到与Itanium2处理器相当的性能。
。     

Efficient Parallel Nonlinear Multigrid Relaxation Algorithms for Low-Level Vision Applications

Multigrid techniques have been shown to significantly improve the convergence rate of the nonlinear relaxation algorithms used in computer vision for the extraction of low-level image features. It is also well known that the computations involved with relaxation algorithms are regular and local, and lead naturally to massive data parallelism. However, standard data parallelism does not exploit the large computing resources of the now available massively parallel 2D processor arrays when coarse image resolutions (i.e., small image grids) have to be processed, like in multigrid methods. In this research note, we present an algorithmic framework which enables us making a full use of the large potential of data parallelism for the implementation of nonlinear multigrid relaxation methods. The approach combines two different levels of parallelism: parallel updating of the image sites and concurrent explorations of the configuration space of the problem. The efficiency of the method is demonstrated on two different low-level vision applications: restoration of noisy images and optical flow computation.     

Kokkos: Enabling manycore performance portability through polymorphic memory access patterns

The manycore revolution can be characterized by increasing thread counts, decreasing memory per thread, and diversity of continually evolving manycore architectures. High performance computing (HPC) applications and libraries must exploit increasingly finer levels of parallelism within their codes to sustain scalability on these devices. A major obstacle to performance portability is the diverse and conflicting set of constraints on memory access patterns across devices. Contemporary portable programming models address manycore parallelism (e.g., OpenMP, OpenACC, OpenCL) but fail to address memory access patterns. The Kokkos C++ library enables applications and domain libraries to achieve performance portability on diverse manycore architectures by unifying abstractions for both fine-grain data parallelism and memory access patterns. In this paper we describe Kokkos’ abstractions, summarize its application programmer interface (API), present performance results for unit-test kernels and mini-applications, and outline an incremental strategy for migrating legacy C++ codes to Kokkos. The Kokkos library is under active research and development to incorporate capabilities from new generations of manycore architectures, and to address a growing list of applications and domain libraries.     

Exploiting Distributed-Memory and Shared-Memory Parallelism on Clusters of SMPs with Data Parallel Programs

Clusters of SMPs are hybrid-parallel architectures that combine the main concepts of distributed-memory and shared-memory parallel machines. Although SMP clusters are widely used in the high performance computing community, there exists no single programming paradigm that allows exploiting the hierarchical structure of these machines. Most parallel applications deployed on SMP clusters are based on MPI, the standard API for distributed-memory parallel programming, and thus may miss a number of optimization opportunities offered by the shared memory available within SMP nodes. In this paper we present extensions to the data parallel programming language HPF and associated compilation techniques for optimizing HPF programs on clusters of SMPs. The proposed extensions enable programmers to control key aspects of distributed-memory and shared-memory parallelization at a high-level of abstraction. Based on these language extensions, a compiler can adopt a hybrid parallelization strategy which closely reflects the hierarchical structure of SMP clusters by automatically exploiting shared-memory parallelism based on OpenMP within cluster nodes and distributed-memory parallelism utilizing MPI across nodes. We describe the implementation of these features in the VFC compiler and present experimental results which show the effectiveness of these techniques.     

Bridging the processor-memory performance gap with 3D IC technology

Microprocessor performance has been improving at roughly 60% per year. Memory access times, however, have improved by less than 10% per year. The resulting gap between logic and memory performance has forced microprocessor designs toward complex and power-hungry architectures that support out-of-order and speculative execution. Moreover, processors have been designed with increasingly large cache hierarchies to hide main memory latency. This article examines how 3D IC technology can improve interactions between the processor and memory. Our work examines the performance of a single-core, single-threaded processor under representative work loads. We have shown that reducing memory latency by bringing main memory on chip gives us near-perfect performance. Three-dimensional IC technology can provide the much needed bandwidth without the cost, design complexity, and power issues associated with a large number of off-chip pins. The principal challenge remains the demonstration of a highly manufacturable 3D IC technology with high yield and low cost.     

Retargeting sequential image-processing programs for data parallel execution

New compact, low-power implementation technologies for processors and imaging arrays can enable a new generation of portable video products. However, software compatibility with large bodies of existing applications written in C prevents more efficient, higher performance data parallel architectures from being used in these embedded products. If this software could be automatically retargeted explicitly for data parallel execution, product designers could incorporate these architectures into embedded products. The key challenge is exposing the parallelism that is inherent in these applications but that is obscured by artifacts imposed by sequential programming languages. This paper presents a recognition-based approach for automatically extracting a data parallel program model from sequential image processing code and retargeting it to data parallel execution mechanisms. The explicitly parallel model presented, called multidimensional data flow (MDDF), captures a model of how operations on data regions (e.g., rows, columns, and tiled blocks) are composed and interact. To extract an MDDF model, a partial recognition technique is used that focuses on identifying array access patterns in loops, transforming only those program elements that hinder parallelization, while leaving the core algorithmic computations intact. The paper presents results of retargeting a set of production programs to a representative data parallel processor array to demonstrate the capacity to extract parallelism using this technique. The retargeted applications yield a potential execution throughput limited only by the number of processing elements, exceeding thousands of instructions per cycle in massively parallel implementations.     

Efficient multimedia coprocessor with enhanced SIMD engines for exploiting ILP and DLP

Multimedia applications have become increasingly important in daily computing. These applications are composed of heterogeneous regions of code mixed with data-level parallelism (DLP) and instruction-level parallelism (ILP). A standard solution for a multimedia coprocessor resembles of single-instruction multiple-data (SIMD) engines into architectures exploiting ILP at compile time, such as very long instruction word (VLIW) and transport triggered architecture (TTA). However, the ILP regions fail to scale with the increased vector length to achieve high performance in the DLP regions. Furthermore, the register-to-register nature of SIMD instructions causes current SIMD engines to have limitations in handling memory alignment, data reorganization, and control flow. Many supporting instructions such as data permutations, address generations, and loop branches, are required to aid in the execution of the real SIMD computation instructions. To mitigate these problems, we propose optimized SIMD engines that have the capabilities for combining VLIW or TTA processing with a unified scalar and long vector computations as well as efficient SIMD hardware for real computation. Our new architecture is based on TTA and is called multimedia coprocessor (MCP). This architecture includes following features: (1) a simple coprocessor structure with 8-way TTA, (2) cost-effective SIMD hardware capable of performing floating-point operations, (3) long vector capabilities built upon existing SIMD hardware and a single register file and processor data path for both scalar operands and vector elements, and (4) an optimized SIMD architecture that addresses the SIMD limitations. Our experimental evaluations show that MCP can outperform conventional SIMD techniques by an average of 39% and 12% in performance for multimedia kernels and applications, respectively.     

Pc-based Shared Memory Architecture and Language

The Image Processing applications require both computing and communication power. The object of the GFLOPS project was to study all aspects concerning the design of such computers. The project's aim was to develop a parallel architecture as well as its software environment to implement these applications efficiently. A development environment, especially a C data-parallel language, has been built for this purpose. The C parallel language presented here, simplifies the use of such architectures by providing the programmer with a global name space and a control mechanism to exploit fine and medium grain parallelism of its applications. The main advantage of our paradigm is that it allows a unique framework to express both data and control parallelism. We have implemented this programming environment on the GFLOPS machine which supports up to 512 processor nodes, which are PC mother boards, connected over a scaleable and cost-effective network, via the PCI-bus, at a constant cost per node. The aim is to obtain at low cost a scaleable virtually shared memory machine. In this paper we discuss the design of the GFLOPS machine and its C parallel language, and evaluate the effectiveness of the mechanisms incorporated. The analysis of the architecture's behaviour was conducted with microbenchmarks and image processing algorithms, written in C.     

The vector-thread architecture

The vector-thread (VT) architectural paradigm describes a class of architectures that unify the vector and multithreaded execution models. VT architectures compactly encode large amounts of structured parallelism in a form that lets simple microarchitectures attain high performance at low power by avoiding complex control and data path structures and by reducing activity on long wires     

异构多核上多级并行模型支持及性能优化

低功耗及廉价性使得异构多核在超级计算机计算资源中占有重要比例.然而,异构多核具有高带宽及松耦合一致性等特点,获得理想的存储及计算性能需要更多地考虑底层硬件细节.实现了一种针对典型的异构多核Cell BE 处理器的多级并行模型CellMLP,通过C 语言扩展编译指导语句,实现了对数据并行、任务并行以及流水并行编程模型的支持,提高了并行程序生产率.运行支持优化方面,数据并行采用SPE 并行数据传输、双缓冲等优化手段来提高数据传输带宽;任务并行使用一种新式混合任务队列以支持异步任务窃取,降低SPE 线程间竞争,提高了任务并行的可扩展性;流水并行首次使用阻塞信号传输机制实现SPE 线程间的低开销同步操作.实验对Stream,NASBenchmark 及BOTS 等应用进行了测试,结果表明,CellMLP 可对多种典型并行应用进行高效支持.与目前同类编程模型SARC 及CellSs 进行性能对比,其结果表明,CellMLP 实际数据传输带宽以及非规则应用的支持方面具有明显优势.     

The Shared Control Parallel Architecture Model

SIMD machines are considered special purpose architectures chiefly because of their inability to support control parallelism. This restriction exists because there is a single control unit that is shared at the thread level; concurrent control threads must time-share the control unit (they are sequentially executed). We present an alternative model for building centralized control architectures that allows better support for control parallelism. This model, called shared control, shares the control unit(s) at the instruction level. More precisely, in each cycle the control signals for all the supported instructions are broadcast to the PEs. In turn, each PE receives its control by synchronizing with the control unit responsible for its local instruction. The shared control model is fundamentally different from the SIMD model. There are a number of architectural issues that must be resolved in order for this model to be useful. This paper identifies some of these issues and discusses their respective trade-off spaces. An integrated shared-control/SIMD architecture design (SharC) is presented and used to demonstrate the relative performance relative to a SIMD architecture.