Supercomputing's Exaflop Target
Total Page:16
File Type:pdf, Size:1020Kb
Load more
Recommended publications
-
Linpack Evaluation on a Supercomputer with Heterogeneous Accelerators
Linpack Evaluation on a Supercomputer with Heterogeneous Accelerators Toshio Endo Akira Nukada Graduate School of Information Science and Engineering Global Scientific Information and Computing Center Tokyo Institute of Technology Tokyo Institute of Technology Tokyo, Japan Tokyo, Japan [email protected] [email protected] Satoshi Matsuoka Naoya Maruyama Global Scientific Information and Computing Center Global Scientific Information and Computing Center Tokyo Institute of Technology/National Institute of Informatics Tokyo Institute of Technology Tokyo, Japan Tokyo, Japan [email protected] [email protected] Abstract—We report Linpack benchmark results on the Roadrunner or other systems described above, it includes TSUBAME supercomputer, a large scale heterogeneous system two types of accelerators. This is due to incremental upgrade equipped with NVIDIA Tesla GPUs and ClearSpeed SIMD of the system, which has been the case in commodity CPU accelerators. With all of 10,480 Opteron cores, 640 Xeon cores, 648 ClearSpeed accelerators and 624 NVIDIA Tesla GPUs, clusters; they may have processors with different speeds as we have achieved 87.01TFlops, which is the third record as a result of incremental upgrade. In this paper, we present a heterogeneous system in the world. This paper describes a Linpack implementation and evaluation results on TSUB- careful tuning and load balancing method required to achieve AME with 10,480 Opteron cores, 624 Tesla GPUs and 648 this performance. On the other hand, since the peak speed is ClearSpeed accelerators. In the evaluation, we also used a 163 TFlops, the efficiency is 53%, which is lower than other systems. -
2020 ALCF Science Report
ARGONNE LEADERSHIP 2020 COMPUTING FACILITY Science On the cover: A snapshot of a visualization of the SARS-CoV-2 viral envelope comprising 305 million atoms. A multi-institutional research team used multiple supercomputing resources, including the ALCF’s Theta system, to optimize codes in preparation for large-scale simulations of the SARS-CoV-2 spike protein that were recognized with the ACM Gordon Bell Special Prize for HPC-Based COVID-19 Research. Image: Rommie Amaro, Lorenzo Casalino, Abigail Dommer, and Zied Gaieb, University of California San Diego 2020 SCIENCE CONTENTS 03 Message from ALCF Leadership 04 Argonne Leadership Computing Facility 10 Advancing Science with HPC 06 About ALCF 12 ALCF Resources Contribute to Fight Against COVID-19 07 ALCF Team 16 Edge Services Propel Data-Driven Science 08 ALCF Computing Resources 18 Preparing for Science in the Exascale Era 26 Science 28 Accessing ALCF GPCNeT: Designing a Benchmark 43 Materials Science 51 Physics Resources for Science Suite for Inducing and Measuring Constructing and Navigating Hadronic Light-by-Light Scattering Contention in HPC Networks Polymorphic Landscapes of and Vacuum Polarization Sudheer Chunduri Molecular Crystals Contributions to the Muon 30 2020 Science Highlights Parallel Relational Algebra for Alexandre Tkatchenko Anomalous Magnetic Moment Thomas Blum 31 Biological Sciences Logical Inferencing at Scale Data-Driven Materials Sidharth Kumar Scalable Reinforcement-Learning- Discovery for Optoelectronic The Last Journey Based Neural Architecture Search Applications -
Tsubame 2.5 Towards 3.0 and Beyond to Exascale
Being Very Green with Tsubame 2.5 towards 3.0 and beyond to Exascale Satoshi Matsuoka Professor Global Scientific Information and Computing (GSIC) Center Tokyo Institute of Technology ACM Fellow / SC13 Tech Program Chair NVIDIA Theater Presentation 2013/11/19 Denver, Colorado TSUBAME2.0 NEC Confidential TSUBAME2.0 Nov. 1, 2010 “The Greenest Production Supercomputer in the World” TSUBAME 2.0 New Development >600TB/s Mem BW 220Tbps NW >12TB/s Mem BW >400GB/s Mem BW >1.6TB/s Mem BW Bisecion BW 80Gbps NW BW 35KW Max 1.4MW Max 32nm 40nm ~1KW max 3 Performance Comparison of CPU vs. GPU 1750 GPU 200 GPU ] 1500 160 1250 GByte/s 1000 120 750 80 500 CPU CPU 250 40 Peak Performance [GFLOPS] Performance Peak 0 Memory Bandwidth [ Bandwidth Memory 0 x5-6 socket-to-socket advantage in both compute and memory bandwidth, Same power (200W GPU vs. 200W CPU+memory+NW+…) NEC Confidential TSUBAME2.0 Compute Node 1.6 Tflops Thin 400GB/s Productized Node Mem BW as HP 80GBps NW ProLiant Infiniband QDR x2 (80Gbps) ~1KW max SL390s HP SL390G7 (Developed for TSUBAME 2.0) GPU: NVIDIA Fermi M2050 x 3 515GFlops, 3GByte memory /GPU CPU: Intel Westmere-EP 2.93GHz x2 (12cores/node) Multi I/O chips, 72 PCI-e (16 x 4 + 4 x 2) lanes --- 3GPUs + 2 IB QDR Memory: 54, 96 GB DDR3-1333 SSD:60GBx2, 120GBx2 Total Perf 2.4PFlops Mem: ~100TB NEC Confidential SSD: ~200TB 4-1 2010: TSUBAME2.0 as No.1 in Japan > All Other Japanese Centers on the Top500 COMBINED 2.3 PetaFlops Total 2.4 Petaflops #4 Top500, Nov. -
Lessons Learned in Deploying the World's Largest Scale Lustre File
Lessons Learned in Deploying the World’s Largest Scale Lustre File System Galen M. Shipman, David A. Dillow, Sarp Oral, Feiyi Wang, Douglas Fuller, Jason Hill, Zhe Zhang Oak Ridge Leadership Computing Facility, Oak Ridge National Laboratory Oak Ridge, TN 37831, USA fgshipman,dillowda,oralhs,fwang2,fullerdj,hilljj,[email protected] Abstract 1 Introduction The Spider system at the Oak Ridge National Labo- The Oak Ridge Leadership Computing Facility ratory’s Leadership Computing Facility (OLCF) is the (OLCF) at Oak Ridge National Laboratory (ORNL) world’s largest scale Lustre parallel file system. Envi- hosts the world’s most powerful supercomputer, sioned as a shared parallel file system capable of de- Jaguar [2, 14, 7], a 2.332 Petaflop/s Cray XT5 [5]. livering both the bandwidth and capacity requirements OLCF also hosts an array of other computational re- of the OLCF’s diverse computational environment, the sources such as a 263 Teraflop/s Cray XT4 [1], visual- project had a number of ambitious goals. To support the ization, and application development platforms. Each of workloads of the OLCF’s diverse computational plat- these systems requires a reliable, high-performance and forms, the aggregate performance and storage capacity scalable file system for data storage. of Spider exceed that of our previously deployed systems Parallel file systems on leadership-class systems have by a factor of 6x - 240 GB/sec, and 17x - 10 Petabytes, traditionally been tightly coupled to single simulation respectively. Furthermore, Spider supports over 26,000 platforms. This approach had resulted in the deploy- clients concurrently accessing the file system, which ex- ment of a dedicated file system for each computational ceeds our previously deployed systems by nearly 4x. -
Titan: a New Leadership Computer for Science
Titan: A New Leadership Computer for Science Presented to: DOE Advanced Scientific Computing Advisory Committee November 1, 2011 Arthur S. Bland OLCF Project Director Office of Science Statement of Mission Need • Increase the computational resources of the Leadership Computing Facilities by 20-40 petaflops • INCITE program is oversubscribed • Programmatic requirements for leadership computing continue to grow • Needed to avoid an unacceptable gap between the needs of the science programs and the available resources • Approved: Raymond Orbach January 9, 2009 • The OLCF-3 project comes out of this requirement 2 ASCAC – Nov. 1, 2011 Arthur Bland INCITE is 2.5 to 3.5 times oversubscribed 2007 2008 2009 2010 2011 2012 3 ASCAC – Nov. 1, 2011 Arthur Bland What is OLCF-3 • The next phase of the Leadership Computing Facility program at ORNL • An upgrade of Jaguar from 2.3 Petaflops (peak) today to between 10 and 20 PF by the end of 2012 with operations in 2013 • Built with Cray’s newest XK6 compute blades • When completed, the new system will be called Titan 4 ASCAC – Nov. 1, 2011 Arthur Bland Cray XK6 Compute Node XK6 Compute Node Characteristics AMD Opteron 6200 “Interlagos” 16 core processor @ 2.2GHz Tesla M2090 “Fermi” @ 665 GF with 6GB GDDR5 memory Host Memory 32GB 1600 MHz DDR3 Gemini High Speed Interconnect Upgradeable to NVIDIA’s next generation “Kepler” processor in 2012 Four compute nodes per XK6 blade. 24 blades per rack 5 ASCAC – Nov. 1, 2011 Arthur Bland ORNL’s “Titan” System • Upgrade of existing Jaguar Cray XT5 • Cray Linux Environment -
Musings RIK FARROWOPINION
Musings RIK FARROWOPINION Rik is the editor of ;login:. While preparing this issue of ;login:, I found myself falling down a rabbit hole, like [email protected] Alice in Wonderland . And when I hit bottom, all I could do was look around and puzzle about what I discovered there . My adventures started with a casual com- ment, made by an ex-Cray Research employee, about the design of current super- computers . He told me that today’s supercomputers cannot perform some of the tasks that they are designed for, and used weather forecasting as his example . I was stunned . Could this be true? Or was I just being dragged down some fictional rabbit hole? I decided to learn more about supercomputer history . Supercomputers It is humbling to learn about the early history of computer design . Things we take for granted, such as pipelining instructions and vector processing, were impor- tant inventions in the 1970s . The first supercomputers were built from discrete components—that is, transistors soldered to circuit boards—and had clock speeds in the tens of nanoseconds . To put that in real terms, the Control Data Corpora- tion’s (CDC) 7600 had a clock cycle of 27 .5 ns, or in today’s terms, 36 4. MHz . This was CDC’s second supercomputer (the 6600 was first), but included instruction pipelining, an invention of Seymour Cray . The CDC 7600 peaked at 36 MFLOPS, but generally got 10 MFLOPS with carefully tuned code . The other cool thing about the CDC 7600 was that it broke down at least once a day . -
Biology at the Exascale
Biology at the Exascale Advances in computational hardware and algorithms that have transformed areas of physics and engineering have recently brought similar benefits to biology and biomedical research. Contributors: Laura Wolf and Dr. Gail W. Pieper, Argonne National Laboratory Biological sciences are undergoing a revolution. High‐performance computing has accelerated the transition from hypothesis‐driven to design‐driven research at all scales, and computational simulation of biological systems is now driving the direction of biological experimentation and the generation of insights. As recently as ten years ago, success in predicting how proteins assume their intricate three‐dimensional forms was considered highly unlikely if there was no related protein of known structure. For those proteins whose sequence resembles a protein of known structure, the three‐dimensional structure of the known protein can be used as a “template” to deduce the unknown protein structure. At the time, about 60 percent of protein sequences arising from the genome sequencing projects had no homologs of known structure. In 2001, Rosetta, a computational technique developed by Dr. David Baker and colleagues at the Howard Hughes Medical Institute, successfully predicted the three‐dimensional structure of a folded protein from its linear sequence of amino acids. (Baker now develops tools to enable researchers to test new protein scaffolds, examine additional structural hypothesis regarding determinants of binding, and ultimately design proteins that tightly bind endogenous cellular proteins.) Two years later, a thirteen‐year project to sequence the human genome was declared a success, making available to scientists worldwide the billions of letters of DNA to conduct postgenomic research, including annotating the human genome. -
(Intel® OPA) for Tsubame 3
CASE STUDY High Performance Computing (HPC) with Intel® Omni-Path Architecture Tokyo Institute of Technology Chooses Intel® Omni-Path Architecture for Tsubame 3 Price/performance, thermal stability, and adaptive routing are key features for enabling #1 on Green 500 list Challenge How do you make a good thing better? Professor Satoshi Matsuoka of the Tokyo Institute of Technology (Tokyo Tech) has been designing and building high- performance computing (HPC) clusters for 20 years. Among the systems he and his team at Tokyo Tech have architected, Tsubame 1 (2006) and Tsubame 2 (2010) have shown him the importance of heterogeneous HPC systems for scientific research, analytics, and artificial intelligence (AI). Tsubame 2, built on Intel® Xeon® Tsubame at a glance processors and Nvidia* GPUs with InfiniBand* QDR, was Japan’s first peta-scale • Tsubame 3, the second- HPC production system that achieved #4 on the Top500, was the #1 Green 500 generation large, production production supercomputer, and was the fastest supercomputer in Japan at the time. cluster based on heterogeneous computing at Tokyo Institute of Technology (Tokyo Tech); #61 on June 2017 Top 500 list and #1 on June 2017 Green 500 list For Matsuoka, the next-generation machine needed to take all the goodness of Tsubame 2, enhance it with new technologies to not only advance all the current • The system based upon HPE and latest generations of simulation codes, but also drive the latest application Apollo* 8600 blades, which targets—which included deep learning/machine learning, AI, and very big data are smaller than a 1U server, analytics—and make it more efficient that its predecessor. -
Germany's Top500 Businesses
GERMANY’S TOP500 DIGITAL BUSINESS MODELS IN SEARCH OF BUSINESS CONTENTS FOREWORD 3 INSIGHT 1 4 SLOW GROWTH RATES YET HIGH SALES INSIGHT 2 6 NOT ENOUGH REVENUE IS ATTRIBUTABLE TO DIGITIZATION INSIGHT 3 10 EU REGULATIONS ARE WEAKENING INNOVATION INSIGHT 4 12 THE GERMAN FEDERAL GOVERNMENT COULD TURN INTO AN INNOVATION DRIVER CONCLUSION 14 FOREWORD Large German companies are on the lookout. Their purpose: To find new growth prospects. While revenue increases of more than 5 percent on average have not been uncommon for Germany’s 500 largest companies in the past, that level of growth has not occurred for the last four years. The reasons are obvious. With their high export rates, This study is intended to examine critically the Germany’s industrial companies continue to be major opportunities arising at the beginning of a new era of players in the global market. Their exports have, in fact, technology. Accenture uses four insights to not only been so high in the past that it is now increasingly describe the progress that has been made in digital difficult to sustain their previous rates of growth. markets, but also suggests possible steps companies can take to overcome weak growth. Accenture regularly examines Germany’s largest companies on the basis of their ranking in “Germany’s Top500,” a list published every year in the German The four insights in detail: daily newspaper DIE WELT. These 500 most successful German companies generate revenue of more than INSIGHT 1 one billion Euros. In previous years, they were the Despite high levels of sales, growth among Germany’s engines of the German economy. -
Jaguar Supercomputer
Jaguar Supercomputer Jake Baskin '10, Jānis Lībeks '10 Jan 27, 2010 What is it? Currently the fastest supercomputer in the world, at up to 2.33 PFLOPS, located at Oak Ridge National Laboratory (ORNL). Leader in "petascale scientific supercomputing". Uses Massively parallel simulations. Modeling: Climate Supernovas Volcanoes Cellulose http://www.nccs.gov/wp- content/themes/nightfall/img/jaguarXT5/gallery/jaguar-1.jpg Overview Processor Specifics Network Architecture Programming Models NCCS networking Spider file system Scalability The guts 84 XT4 and 200 XT5 cabinets XT5 18688 compute nodes 256 service and i/o nodes XT4 7832 compute nodes 116 service and i/o nodes (XT5) Compute Nodes 2 Opteron 2435 Istanbul (6 core) processors per node 64K L1 instruction cache 65K L1 data cache per core 512KB L2 cache per core 6MB L3 cache per processor (shared) 8GB of DDR2-800 RAM directly attached to each processor by integrated memory controller. http://www.cray.com/Assets/PDF/products/xt/CrayXT5Brochure.pdf How are they organized? 3-D Torus topology XT5 and XT4 segments are connected by an InfiniBand DDR network 889 GB/sec bisectional bandwidth http://www.cray.com/Assets/PDF/products/xt/CrayXT5Brochure.pdf Programming Models Jaguar supports these programming models: MPI (Message Passing Interface) OpenMP (Open Multi Processing) SHMEM (SHared MEMory access library) PGAS (Partitioned global address space) NCCS networking Jaguar usually performs computations on large datasets. These datasets have to be transferred to ORNL. Jaguar is connected to ESnet (Energy Sciences Network, scientific institutions) and Internet2 (higher education institutions). ORNL owns its own optical network that allows 10Gb/s to various locations around the US. -
ECP Software Technology Capability Assessment Report
ECP-RPT-ST-0001-2018 ECP Software Technology Capability Assessment Report Michael A. Heroux, Director ECP ST Jonathan Carter, Deputy Director ECP ST Rajeev Thakur, Programming Models & Runtimes Lead Jeffrey Vetter, Development Tools Lead Lois Curfman McInnes, Mathematical Libraries Lead James Ahrens, Data & Visualization Lead J. Robert Neely, Software Ecosystem & Delivery Lead July 1, 2018 DOCUMENT AVAILABILITY Reports produced after January 1, 1996, are generally available free via US Department of Energy (DOE) SciTech Connect. Website http://www.osti.gov/scitech/ Reports produced before January 1, 1996, may be purchased by members of the public from the following source: National Technical Information Service 5285 Port Royal Road Springfield, VA 22161 Telephone 703-605-6000 (1-800-553-6847) TDD 703-487-4639 Fax 703-605-6900 E-mail [email protected] Website http://www.ntis.gov/help/ordermethods.aspx Reports are available to DOE employees, DOE contractors, Energy Technology Data Exchange representatives, and International Nuclear Information System representatives from the following source: Office of Scientific and Technical Information PO Box 62 Oak Ridge, TN 37831 Telephone 865-576-8401 Fax 865-576-5728 E-mail [email protected] Website http://www.osti.gov/contact.html This report was prepared as an account of work sponsored by an agency of the United States Government. Neither the United States Government nor any agency thereof, nor any of their employees, makes any warranty, express or implied, or assumes any legal liability or responsibility for the accuracy, completeness, or usefulness of any information, apparatus, product, or process disclosed, or represents that its use would not infringe privately owned rights. -
Experimental and Analytical Study of Xeon Phi Reliability
View metadata, citation and similar papers at core.ac.uk brought to you by CORE provided by Lume 5.8 Experimental and Analytical Study of Xeon Phi Reliability Daniel Oliveira Laércio Pilla Nathan DeBardeleben Institute of Informatics, UFRGS Department of Informatics and Los Alamos National Laboratory Porto Alegre, RS, Brazil Statistics, UFSC Los Alamos, NM, US Florianópolis, SC, Brazil Sean Blanchard Heather Quinn Israel Koren Los Alamos National Laboratory Los Alamos National Laboratory University of Massachusetts, UMass Los Alamos, NM, US Los Alamos, NM, US Amherst, MA, US Philippe Navaux Paolo Rech Institute of Informatics, UFRGS Institute of Informatics, UFRGS Porto Alegre, RS, Brazil Porto Alegre, RS, Brazil ABSTRACT 1 INTRODUCTION We present an in-depth analysis of transient faults effects on HPC Accelerators are extensively used to expedite calculations in large applications in Intel Xeon Phi processors based on radiation experi- HPC centers. Tianhe-2, Cori, Trinity, and Oakforest-PACS use Intel ments and high-level fault injection. Besides measuring the realistic Xeon Phi and many other top supercomputers use other forms of error rates of Xeon Phi, we quantify Silent Data Corruption (SDCs) accelerators [17]. The main reasons to use accelerators are their by correlating the distribution of corrupted elements in the out- high computational capacity, low cost, reduced per-task energy put to the application’s characteristics. We evaluate the benefits consumption, and flexible development platforms. Unfortunately, of imprecise computing for reducing the programs’ error rate. For accelerators are also extremely likely to experience transient errors example, for HotSpot a 0.5% tolerance in the output value reduces as they are built with cutting-edge technology, have very high the error rate by 85%.