Heterogeneous Computing in the Cloud: Crunching Big Data and Democratizing HPC Access for the Life Sciences

Total Page:16

File Type:pdf, Size:1020Kb

Heterogeneous Computing in the Cloud: Crunching Big Data and Democratizing HPC Access for the Life Sciences WHITE PAPER Intel® Xeon® Processors Intel® Xeon Phi™ Coprocessor Life Sciences Computing Heterogeneous Computing in the Cloud: Crunching Big Data and Democratizing HPC Access for the Life Sciences A New Paradigm for Life Sciences Computing The combination of heterogeneous computing and cloud computing is emerging as a powerful new paradigm to meet the requirements for high-performance computing (HPC) and data throughput throughout the life sciences (LS) and healthcare value chains. Of course, neither cloud computing nor the use of innovative computing architectures is new, but the rise of big data as a defining feature of modern life sciences and the proliferation of vastly differing applications to mine the data have dramatically changed the landscape of LS computing requirements. Heterogeneous cloud computing offers heterogeneous architectures can the potential to shift flexibly from one provide even small research groups HPC architecture to another in a secure with affordable access to diverse public or private cloud environment. As compute resources. such, it meets two critical needs for life This paper discusses several trends and sciences computing: enablers of affordable, heterogeneous • Crunch more data faster. As data cloud computing for LS, including the new sets have ballooned, HPC approaches Intel® Xeon Phi™ coprocessor, based on Intel® have evolved and diversified to better Many Integrated Core Architecture (Intel® match specific problems with their most MIC Architecture). The Intel Xeon Phi effective HPC solutions. Indeed, some LS coprocessor represents a breakthrough in problems (for example, de novo genome heterogeneous computing by delivering assembly) are essentially intractable exceptional throughput and energy unless tackled with special-purpose efficiency without the high costs, solutions. Today, no single approach is inflexibility, and programming challenges adequate. Heterogeneous computing that have plagued many previous embodies the use of multiple approaches approaches to heterogeneous computing. to achieve optimal throughput for each This paper also provides brief examples big data LS workload. of heterogeneous computing innovators, such as Nimbix and Convey Computer, Ketan Paranjape • Democratize access. While the scope that are adopting or experimenting with Director, Healthcare and complexity of HPC resources have heterogeneous computing approaches. and Life Sciences, grown, the ability of research groups Intel Corporation to identify, afford, and support them Rising Demand for Big Data has diminished. Budget constraints are Steve Hebert Computing CEO, limiting access to necessary compute Nimbix resources at the very time when the Genomics is perhaps the clearest LS example where progress is accelerated Bob Masson explosive growth in LS data makes access increasingly desirable. HPC- by access to the right HPC resource Marketing Manager, and stymied when those resources Convey Computer oriented clouds supporting the latest Heterogeneous Computing in the Cloud: Crunching Big Data and Democratizing HPC Access for the Life Sciences Table of Contents are missing. Pioneering computational Cloud-based, heterogeneous computing biologist Dr. Eric Schadt, currently represents a significant step toward solving A New Paradigm for director of the Institute for Genomics and these problems. Indeed, heterogeneous Life Sciences Computing. 1 Multiscale Biology at Mt. Sinai Hospital, computing has become virtually a necessity Rising Demand for captured the challenge well in a recent in life sciences, where the output from next- Big Data Computing. .1 Nature Reviews Genetics paper. Dr. Schadt generation sequencing (NGS) instruments Application-Specific Computing. 3 and his colleagues wrote: represents a data tipping point. This data deluge has outpaced even the steady Intel Xeon Phi Coprocessor “The astonishing rate of data generation performance-doubling of Moore’s Law. New Advances Performance, by these low-cost, high-throughput Energy, and Programmability. 3 approaches based on specialized processors technologies in genomics is being matched such as field-programmable gate arrays Embracing Heterogeneous by that of other technologies, such as (FPGAs) and general-purpose computing on Computing: Jackson Laboratory. 4 real-time imaging and mass spectrometry- graphics processing units (GPGPUs), as well Other Advances: Improving based flow cytometry. Success in the as innovative computing strategies such as Ease of Use and Development life sciences will depend on our ability Apache Hadoop*, are being pressed into for Heterogeneous Computing. 5 to properly interpret the large-scale, service with impressive results. Democratizing HPC Access high-dimensional data sets that are with Cloud Computing. 6 generated by these technologies, which Not surprisingly, heterogeneous in turn requires us to adopt advances in approaches are being embraced by the Hiding Complexity from the User. .7 informatics…[Scientists must] master bioinformatics community. “We just don’t Entering a New Era. .8 different types of computational have enough electricity, cooling, floor Learn More. .8 environments that exist—such as cloud space, money, etc. for using standard and heterogeneous computing—to clusters or parallel processing to handle successfully tackle our big data problems. the load,” said Dr. Harold “Skip” Garner, head of the Medical Informatics Systems “In under a year genomics technologies Division and former executive director will enable individual laboratories to of the Virginia Bioinformatics Institute, generate terabyte or even petabyte scales both part of the Virginia Polytechnic of data at a reasonable cost. However, Institute and State University (Virginia the computational infrastructure that is Tech). “Future bioinformatics centers, required to maintain and process these especially large computing centers, will large-scale data sets, and to integrate them have a mix of technologies in hardware with other large-scale sets, is typically that include standard processors, FPGAs, beyond the reach of small laboratories and and GPGPUs, and new jobs will be designed is increasingly posing challenges even for for, implemented on, and steered to the large institutes.”1 appropriate processing environments.”2 Figure 1: The Convey Intel® Xeon® Processor-Based Host FPGA-Based Coprocessor Computer hybrid- core system provides a reconfigurable, Xilinx Xilinx Xilinx Xilinx FPGA* FPGA FPGA FPGA application-specific Intel® Xeon® HCMI accelerator, but Processor(s) appears as a standard Intel I/O Application-Specific Personalities Intel® architecture- Subsystem based server to the rest of the computing Cluster Interconnect infrastructure Fabric Memory Memory Memory Hybrid-Core Globally Shared Memory (HCGSM) 2 Heterogeneous Computing in the Cloud: Crunching Big Data and Democratizing HPC Access for the Life Sciences Application-Specific Computing do searching and alignment. Such It is still early days for heterogeneous applications rely on many independent HPC cloud computing. Advances in and simple operations, and are thus virtualization technology, the porting highly parallelizable. of key bioinformatics algorithms to Unique architectures from innovative special-purpose hardware, and ongoing companies are emerging to help meet evolution in standard microprocessor the demand. For example, Convey architecture, such as the Intel Xeon Phi Computer, established in December coprocessor, are all important enablers of 2006, has created a hybrid-core system cloud-based, heterogeneous computing. that pairs classic Intel® processors with While challenges remain (faster data a coprocessor of FPGAs (see Figure 1). transfer and data security concerns, for Particular algorithms—DNA sequence example), heterogeneous HPC in the cloud assembly, for example—are optimized is demonstrating the potential to broadly and translated into code that’s loaded enable researchers and clinicians. onto the FPGAs at runtime. The Convey GPU-based acceleration was an early architecture also features a highly parallel success. First developed to speed graphics memory subsystem to further increase performance in gaming, GPUs’ excellent performance. Convey Computer’s approach floating-point capability proved attractive provides very fast access to random in many applications. GPGPU-based access to memory or single-word access to systems were quickly adopted by the oil memory, and is very useful for the hashing and gas industry, where they delivered functions so widely used in bioinformatics. dramatic speedups at reduced cost. Today, general-purpose GPU computing has Intel Xeon Phi Coprocessor spread to LS disciplines where floating- Advances Performance, Energy, point performance is important (molecular and Programmability modeling, for example, and some Lively debate persists over the right HPC alignment algorithms). approach for various applications and disciplines. For example, FPGA advocates More recently, systems based on FPGAs contend that if you lay out the gates on have gained momentum throughout a chip to execute the specific aspect of genomics and life sciences. While it’s an algorithm, it will by definition be more possible to use application-specific efficient than a series of instructions on a integrated circuits (ASICs), LS applications commodity processor. However, if getting change so quickly that it’s impractical to an algorithm onto a new architecture spin up an ASIC for each algorithm for requires that you use a new language, then every HPC application. This approach
Recommended publications
  • On Heterogeneous Compute and Memory Systems
    ON HETEROGENEOUS COMPUTE AND MEMORY SYSTEMS by Jason Lowe-Power A dissertation submitted in partial fulfillment of the requirements for the degree of Doctor of Philosophy (Computer Sciences) at the UNIVERSITY OF WISCONSIN–MADISON 2017 Date of final oral examination: 05/31/2017 The dissertation is approved by the following members of the Final Oral Committee: Mark D. Hill, Professor, Computer Sciences Dan Negrut, Professor, Mechanical Engineering Jignesh M. Patel, Professor, Computer Sciences Karthikeyan Sankaralingam, Associate Professor, Computer Sciences David A. Wood, Professor, Computer Sciences © Copyright by Jason Lowe-Power 2017 All Rights Reserved i Acknowledgments I would like to acknowledge all of the people who helped me along the way to completing this dissertation. First, I would like to thank my advisors, Mark Hill and David Wood. Often, when students have multiple advisors they find there is high “synchronization overhead” between the advisors. However, Mark and David complement each other well. Mark is a high-level thinker, focusing on the structure of the argument and distilling ideas to their essentials; David loves diving into the details of microarchitectural mechanisms. Although ever busy, at least one of Mark or David were available to meet with me, and they always took the time to help when I needed it. Together, Mark and David taught me how to be a researcher, and they have given me a great foundation to build my career. I thank my committee members. Jignesh Patel for his collaborations, and for the fact that each time I walked out of his office after talking to him, I felt a unique excitement about my research.
    [Show full text]
  • Introduction to Intel Xeon Phi Programming Models
    Introduction to Intel Xeon Phi programming models F.Affinito F. Salvadore SCAI - CINECA Part I Introduction to the Intel Xeon Phi architecture Trends: transistors Trends: clock rates Trends: cores and threads Trends: summarizing... The number of transistors increases The power consumption must not increase The density cannot increase on a single chip Solution : Increase the number of cores GP-GPU and Intel Xeon Phi.. Coupled to the CPU To accelerate highly parallel kernels, facing with the Amdahl Law What is Intel Xeon Phi? 7100 / 5100 / 3100 Series available 5110P: Intel Xeon Phi clock: 1053 MHz 60 cores in-order ~ 1 TFlops/s DP peak performance (2 Tflops SP) 4 hardware threads per core 8 GB DDR5 memory 512-bit SIMD vectors (32 registers) Fully-coherent L1 and L2 caches PCIe bus (rev. 2.0) Max Memory bandwidth (theoretical) 320 GB/s Max TDP: 225 W MIC vs GPU naïve comparison The comparison is naïve System K20s 5110P # cores 2496 60 (*4) Memory size 5 GB 8 GB Peak performance 3.52 TFlops 2 TFlops (SP) Peak performance 1.17 TFlops 1 TFlops (DP) Clock rate 0.706 GHz 1.053 GHz Memory bandwidth 208 GB/s (ECC off) 320 GB/s Terminology MIC = Many Integrated Cores is the name of the architecture Xeon Phi = Commercial name of the Intel product based on the MIC architecture Knight's corner, Knight's landing, Knight's ferry are development names of MIC architectures We will often refer to the CPU as HOST and Xeon Phi as DEVICE Is it an accelerator? YES: It can be used to “accelerate” hot-spots of the code that are highly parallel and computationally extensive In this sense, it works alongside the CPU It can be used as an accelerator using the “offload” programming model An important bottleneck is represented by the communication between host and device (through PCIe) Under this respect, it is very similar to a GPU Is it an accelerator? / 2 NOT ONLY: the Intel Xeon Phi can behave as a many-core X86 node.
    [Show full text]
  • Accelerators for HP Proliant Servers Enable Scalable and Efficient High-Performance Computing
    Family data sheet Accelerators for HP ProLiant servers Enable scalable and efficient high-performance computing November 2014 Family data sheet | Accelerators for HP ProLiant servers HP high-performance computing has made it possible to accelerate innovation at any scale. But traditional CPU technology is no longer capable of sufficiently scaling performance to address the skyrocketing demand for compute resources. HP high-performance computing solutions are built on HP ProLiant servers using industry-leading accelerators to dramatically increase performance with lower power requirements. Innovation is the foundation for success What is hybrid computing? Accelerators are revolutionizing high performance computing A hybrid computing model is one where High-performance computing (HPC) is being used to address many of modern society’s biggest accelerators (known as GPUs or coprocessors) challenges, such as designing new vaccines and genetically engineering drugs to fight diseases, work together with CPUs to perform computing finding and extracting precious oil and gas resources, improving financial instruments, and tasks. designing more fuel efficient engines. As parallel processors, accelerators can split computations into hundreds or thousands of This rapid pace of innovation has created an insatiable demand for compute power. At the same pieces and calculate them simultaneously. time, multiple strict requirements are placed on system performance, power consumption, size, response, reliability, portability, and design time. Modern HPC systems are rapidly evolving, Offloading the most compute-intensive portions of already reaching petaflop and targeting exaflop performance. applications to accelerators dramatically increases both application performance and computational All of these challenges lead to a common set of requirements: a need for more computing efficiency.
    [Show full text]
  • Quick-Reference Guide to Optimization with Intel® Compilers
    Quick Reference Guide to Optimization with Intel® C++ and Fortran Compilers v19.1 For IA-32 processors, Intel® 64 processors, Intel® Xeon Phi™ processors and compatible non-Intel processors. Contents Application Performance .............................................................................................................................. 2 General Optimization Options and Reports ** ............................................................................................. 3 Parallel Performance ** ................................................................................................................................ 4 Recommended Processor-Specific Optimization Options ** ....................................................................... 5 Optimizing for the Intel® Xeon Phi™ x200 product family ............................................................................ 6 Interprocedural Optimization (IPO) and Profile-Guided Optimization (PGO) Options ................................ 7 Fine-Tuning (All Processors) ** ..................................................................................................................... 8 Floating-Point Arithmetic Options .............................................................................................................. 10 Processor Code Name With Instruction Set Extension Name Synonym .................................................... 11 Frequently Used Processor Names in Compiler Options ...........................................................................
    [Show full text]
  • Broadwell Skylake Next Gen* NEW Intel NEW Intel NEW Intel Microarchitecture Microarchitecture Microarchitecture
    15 лет доступности IOTG is extending the product availability for IOTG roadmap products from a minimum of 7 years to a minimum of 15 years when both processor and chipset are on 22nm and newer process technologies. - Xeon Scalable (w/ chipsets) - E3-12xx/15xx v5 and later (w/ chipsets) - 6th gen Core and later (w/ chipsets) - Bay Trail (E3800) and later products (Braswell, N3xxx) - Atom C2xxx (Rangeley) and later - Не включает в себя Xeon-D (7 лет) и E5-26xx v4 (7 лет) 2 IOTG Product Availability Life-Cycle 15 year product availability will start with the following products: Product Discontinuance • Intel® Xeon® Processor Scalable Family codenamed Skylake-SP and later with associated chipsets Notification (PDN)† • Intel® Xeon® E3-12xx/15xx v5 series (Skylake) and later with associated chipsets • 6th Gen Intel® Core™ processor family (Skylake) and later (includes Intel® Pentium® and Celeron® processors) with PDNs will typically be issued no later associated chipsets than 13.5 years after component • Intel Pentium processor N3700 (Braswell) and later and Intel Celeron processors N3xxx (Braswell) and J1900/N2xxx family introduction date. PDNs are (Bay Trail) and later published at https://qdms.intel.com/ • Intel® Atom® processor C2xxx (Rangeley) and E3800 family (Bay Trail) and late Last 7 year product availability Time Last Last Order Ship Last 15 year product availability Time Last Last Order Ship L-1 L L+1 L+2 L+3 L+4 L+5 L+6 L+7 L+8 L+9 L+10 L+11 L+12 L+13 L+14 L+15 Years Introduction of component family † Intel may support this extended manufacturing using reasonably Last Time Order/Ship Periods Component family introduction dates are feasible means deemed by Intel to be appropriate.
    [Show full text]
  • The Intel X86 Microarchitectures Map Version 2.0
    The Intel x86 Microarchitectures Map Version 2.0 P6 (1995, 0.50 to 0.35 μm) 8086 (1978, 3 µm) 80386 (1985, 1.5 to 1 µm) P5 (1993, 0.80 to 0.35 μm) NetBurst (2000 , 180 to 130 nm) Skylake (2015, 14 nm) Alternative Names: i686 Series: Alternative Names: iAPX 386, 386, i386 Alternative Names: Pentium, 80586, 586, i586 Alternative Names: Pentium 4, Pentium IV, P4 Alternative Names: SKL (Desktop and Mobile), SKX (Server) Series: Pentium Pro (used in desktops and servers) • 16-bit data bus: 8086 (iAPX Series: Series: Series: Series: • Variant: Klamath (1997, 0.35 μm) 86) • Desktop/Server: i386DX Desktop/Server: P5, P54C • Desktop: Willamette (180 nm) • Desktop: Desktop 6th Generation Core i5 (Skylake-S and Skylake-H) • Alternative Names: Pentium II, PII • 8-bit data bus: 8088 (iAPX • Desktop lower-performance: i386SX Desktop/Server higher-performance: P54CQS, P54CS • Desktop higher-performance: Northwood Pentium 4 (130 nm), Northwood B Pentium 4 HT (130 nm), • Desktop higher-performance: Desktop 6th Generation Core i7 (Skylake-S and Skylake-H), Desktop 7th Generation Core i7 X (Skylake-X), • Series: Klamath (used in desktops) 88) • Mobile: i386SL, 80376, i386EX, Mobile: P54C, P54LM Northwood C Pentium 4 HT (130 nm), Gallatin (Pentium 4 Extreme Edition 130 nm) Desktop 7th Generation Core i9 X (Skylake-X), Desktop 9th Generation Core i7 X (Skylake-X), Desktop 9th Generation Core i9 X (Skylake-X) • Variant: Deschutes (1998, 0.25 to 0.18 μm) i386CXSA, i386SXSA, i386CXSB Compatibility: Pentium OverDrive • Desktop lower-performance: Willamette-128
    [Show full text]
  • FY18Q2 Marketwide Shopping Event
    $130 BRIGHTON AREA SCHOOL DISTRICT LAPTOP PRESIDENTS DAY EARLY ACCESS STARTS 2/12 DOORBUSTERS ALL WEEK AT 11AM ET 11.6" A DOORBUSTER 11AM ET, 2/16 A: Inspiron 11 3000 Market value* $199.99 | Save $70 $12999 Intel® Celeron® Processor, Windows 10 Home, 2GB memory* & 32GB storage, 11.6" HD display DOORBUSTER 11AM ET, 2/13 23.8" B B: Dell 24 Monitor - S2418NX Market value* $259.99 | Save $100 $15999 23.8" widescreen In nityEdge monitor, C Full HD 1920 x 1080, Dell HDR DOORBUSTER 11AM ET, 2/14 C: Inspiron Small Desktop Market value* $319.99 | Save $100 $21999 Intel® Pentium® processor, Windows 10 Home, 4GB memory* & 1TB* hard drive Limited quantities available at these prices SAVE BIG ON INCREDIBLE COLOR, SOUND AND STREAMING WITH DELL CINEMA PCs. www.Dell.com/BrightonASD US3936579 *See reverse for important details. Netflix streaming membership required. Screen images are simulated. EARLY ACCESS TO DOORBUSTERS PRESIDENTS DAY SALE STARTS EARLY FOR MEMBERS SAVE UP TO 40% A 15.6" B 27" C 15.6" DOORBUSTER 11AM ET, 2/12 DOORBUSTER 11AM ET, 2/15 D A: Inspiron 15 3000 B: New Inspiron Gaming Desktop Market value* $299.99 | Save $70 Market value* $799.99 | Save $100 $22999 $69999 Intel® Celeron® Processor, Windows 10 Home, 8th Gen Intel® Core™ i5 processor, Windows 10 Home, 4GB memory*, 500GB* hard drive, 15.6" HD 8GB memory*, 1TB* hard drive, 3GB* NVIDIA® display GeForce® GTX 1060 graphics card* DOORBUSTER 11AM ET, 2/14 DOORBUSTER 11AM ET, 2/13 C: Dell UltraSharp 27 D: Inspiron 15 5000 2-in-1 4K Monitor - U2718Q Market value* $499.99 | Save $100 Market value* $739.99 | Save $260 $ 99 $ 99 399 479 7th Gen Intel® Core™ i3 processor, 27" 4K In nityEdge monitor, 3840 x 2160 Windows 10 Home, 4GB memory* & 1TB* resolution, Dell HDR hard drive, 15.6" Full HD Touch display *Limited quantities available at these prices.
    [Show full text]
  • Future Computing Platforms for Science in a Power Constrained Era
    Future Computing Platforms for Science in a Power Constrained Era David Abdurachmanov1, Peter Elmer2, Giulio Eulisse1, Robert Knight3 1 Fermilab, Batavia, IL 60510, USA 2 Department of Physics, Princeton University, Princeton, NJ 08540, USA 3 Research Computing, Office of Information Technology, Princeton University, Princeton, NJ, 08540, USA E-mail: [email protected] Abstract. Power consumption will be a key constraint on the future growth of Distributed High Throughput Computing (DHTC) as used by High Energy Physics (HEP). This makes performance-per-watt a crucial metric for selecting cost-efficient computing solutions. For this paper, we have done a wide survey of current and emerging architectures becoming available on the market including x86-64 variants, ARMv7 32-bit, ARMv8 64-bit, Many-Core and GPU solutions, as well as newer System-on-Chip (SoC) solutions. We compare performance and energy efficiency using an evolving set of standardized HEP-related benchmarks and power measurement techniques we have been developing. We evaluate the potential for use of such computing solutions in the context of DHTC systems, such as the Worldwide LHC Computing Grid (WLCG). 1. Introduction and Motivation The data produced by the four experiments at the Large Hadron Collider (LHC) [1] or similar High Energy Physics (HEP) experiments requires a significant amount of human and computing resources which cannot be provided by research institute or even country. For this reasons the various parties involved created the Worldwide LHC Computing Grid (WLCG) in order to tackle the data processing challenges posed by such a large amount of data. The WLGC consists of a arXiv:1510.03676v1 [cs.DC] 28 Jul 2015 highly federated union of computing centers sparse in 40 countries and it represents an admirable example of international organization.
    [Show full text]
  • Introduction to Intel Xeon Phi (“Knights Landing”) on Cori
    Introduction to Intel Xeon Phi (“Knights Landing”) on Cori" Brandon Cook! Brian Friesen" 2017 June 9 - 1 - Knights Landing is here!" • KNL nodes installed in Cori in 2016 • “Pre-produc=on” for ~ 1 year – No charge for using KNL nodes – Limited access (un7l now!) – Intermi=ent down7me – Frequent so@ware changes • KNL nodes on Cori will soon enter produc=on – Charging Begins 2017 July 1 - 2 - Knights Landing overview" Knights Landing: Next Intel® Xeon Phi™ Processor Intel® Many-Core Processor targeted for HPC and Supercomputing First self-boot Intel® Xeon Phi™ processor that is binary compatible with main line IA. Boots standard OS. Significant improvement in scalar and vector performance Integration of Memory on package: innovative memory architecture for high bandwidth and high capacity Integration of Fabric on package Three products KNL Self-Boot KNL Self-Boot w/ Fabric KNL Card (Baseline) (Fabric Integrated) (PCIe-Card) Potential future options subject to change without notice. All timeframes, features, products and dates are preliminary forecasts and subject to change without further notification. - 3 - Knights Landing overview TILE CHA 2 VPU 2 VPU Knights Landing Overview 1MB L2 Core Core 2 x16 X4 MCDRAM MCDRAM 1 x4 DMI MCDRAM MCDRAM Chip: 36 Tiles interconnected by 2D Mesh Tile: 2 Cores + 2 VPU/core + 1 MB L2 EDC EDC PCIe D EDC EDC M Gen 3 3 I 3 Memory: MCDRAM: 16 GB on-package; High BW D Tile D D D DDR4: 6 channels @ 2400 up to 384GB R R 4 36 Tiles 4 IO: 36 lanes PCIe Gen3. 4 lanes of DMI for chipset C DDR MC connected by DDR MC C Node: 1-Socket only H H A 2D Mesh A Fabric: Omni-Path on-package (not shown) N Interconnect N N N E E L L Vector Peak Perf: 3+TF DP and 6+TF SP Flops S S Scalar Perf: ~3x over Knights Corner EDC EDC misc EDC EDC Streams Triad (GB/s): MCDRAM : 400+; DDR: 90+ Source Intel: All products, computer systems, dates and figures specified are preliminary based on current expectations, and are subject to change without notice.
    [Show full text]
  • Heterogeneous Cpu+Gpu Computing
    HETEROGENEOUS CPU+GPU COMPUTING Ana Lucia Varbanescu – University of Amsterdam [email protected] Significant contributions by: Stijn Heldens (U Twente), Jie Shen (NUDT, China), Heterogeneous platforms • Systems combining main processors and accelerators • e.g., CPU + GPU, CPU + Intel MIC, AMD APU, ARM SoC • Everywhere from supercomputers to mobile devices Heterogeneous platforms • Host-accelerator hardware model Accelerator FPGAs Accelerator PCIe / Shared memory ... Host MICs Accelerator GPUs Accelerator CPUs Our focus today … • A heterogeneous platform = CPU + GPU • Most solutions work for other/multiple accelerators • An application workload = an application + its input dataset • Workload partitioning = workload distribution among the processing units of a heterogeneous system Few cores Thousands of Cores 5 Generic multi-core CPU 6 Programming models • Pthreads + intrinsics • TBB – Thread building blocks • Threading library • OpenCL • To be discussed … • OpenMP • Traditional parallel library • High-level, pragma-based • Cilk • Simple divide-and-conquer model abstractionincreasesLevel of 7 A GPU Architecture Offloading model Kernel Host code 9 Programming models • CUDA • NVIDIA proprietary • OpenCL • Open standard, functionally portable across multi-cores • OpenACC • High-level, pragma-based • Different libraries, programming models, and DSLs for different domains Level of abstractionincreasesLevel of CPU vs. GPU 10 ALU ALU CPU Control Low latency, high Throughput: ~ALU 500 GFLOPsALU flexibility. Bandwidth: ~ 60 GB/s Excellent for
    [Show full text]
  • Intel® Omni-Path Architecture Overview and Update
    The architecture for Discovery June, 2016 Intel Confidential Caught in the Vortex…? Business Efficiency & Agility DATA: Trust, Privacy, sovereignty Innovation: New Economy Biz Models Macro Economic Effect Growth Enablers/Inhibitors Intel® Solutions Summit 2016 2 Intel® Solutions Summit 2016 3 Intel Confidential 4 Data Center Blocks Reduce Complexity Intel engineering, validation, support Data Center Blocks Speed time to market Begin with a higher level of integration HPC Cloud Enterprise Storage Increase Value VSAN Ready HPC Compute SMB Server Block Reduce TCO, value pricing Block Node Fuel innovation Server blocks for specific segments Focus R&D on value-add and differentiation Intel® Solutions Summit 2016 5 A Holistic Design Solution for All HPC Needs Intel® Scalable System Framework Small Clusters Through Supercomputers Compute Memory/Storage Compute and Data-Centric Computing Fabric Software Standards-Based Programmability On-Premise and Cloud-Based Intel Silicon Photonics Intel® Xeon® Processors Intel® Solutions for Lustre* Intel® Omni-Path Architecture HPC System Software Stack Intel® Xeon Phi™ Processors Intel® SSDs Intel® True Scale Fabric Intel® Software Tools Intel® Xeon Phi™ Coprocessors Intel® Optane™ Technology Intel® Ethernet Intel® Cluster Ready Program Intel® Server Boards and Platforms 3D XPoint™ Technology Intel® Silicon Photonics Intel® Visualization Toolkit Intel Confidential 14 Parallel is the Path Forward Intel® Xeon® and Intel® Xeon Phi™ Product Families are both going parallel How do we attain extremely high compute
    [Show full text]
  • State-Of-The-Art in Heterogeneous Computing
    Scientific Programming 18 (2010) 1–33 1 DOI 10.3233/SPR-2009-0296 IOS Press State-of-the-art in heterogeneous computing Andre R. Brodtkorb a,∗, Christopher Dyken a, Trond R. Hagen a, Jon M. Hjelmervik a and Olaf O. Storaasli b a SINTEF ICT, Department of Applied Mathematics, Blindern, Oslo, Norway E-mails: {Andre.Brodtkorb, Christopher.Dyken, Trond.R.Hagen, Jon.M.Hjelmervik}@sintef.no b Oak Ridge National Laboratory, Future Technologies Group, Oak Ridge, TN, USA E-mail: [email protected] Abstract. Node level heterogeneous architectures have become attractive during the last decade for several reasons: compared to traditional symmetric CPUs, they offer high peak performance and are energy and/or cost efficient. With the increase of fine-grained parallelism in high-performance computing, as well as the introduction of parallelism in workstations, there is an acute need for a good overview and understanding of these architectures. We give an overview of the state-of-the-art in heterogeneous computing, focusing on three commonly found architectures: the Cell Broadband Engine Architecture, graphics processing units (GPUs), and field programmable gate arrays (FPGAs). We present a review of hardware, available software tools, and an overview of state-of-the-art techniques and algorithms. Furthermore, we present a qualitative and quantitative comparison of the architectures, and give our view on the future of heterogeneous computing. Keywords: Power-efficient architectures, parallel computer architecture, stream or vector architectures, energy and power consumption, microprocessor performance 1. Introduction the speed of logic gates, making computers smaller and more power efficient. Noyce and Kilby indepen- The goal of this article is to provide an overview of dently invented the integrated circuit in 1958, leading node-level heterogeneous computing, including hard- to further reductions in power and space required for ware, software tools and state-of-the-art algorithms.
    [Show full text]