DOCSLIB.ORG

Advanced Architectures Goals of Distributed Computing Evaluating

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
Advanced Architectures Goals of Distributed Computing Evaluating 6/7/2018 Advanced Architectures Goals of Distributed Computing 15A. Distributed Computing • better services 15B. Multi-Processor Systems – scalability • apps too big to run on a single computer 15C. Tightly Coupled Distributed Systems • grow system capacity to meet growing demand 15D. Loosely Coupled Distributed Systems – improved reliability and availability 15E. Cloud Models – improved ease of use, reduced CapEx/OpEx 15F. Distributed Middle-Ware • new services – applications that span multiple system boundaries – global resource domains, services (vs. systems) – complete location transparency Advanced Architectures 1 Advanced Architectures 2 Major Classes of Distributed Systems Evaluating Distributed Systems • Symmetric Multi-Processors (SMP) • Performance – multiple CPUs, sharing memory and I/O devices – overhead, scalability, availability • Single-System Image (SSI) & Cluster Computing • Functionality – a group of computers, acting like a single computer – adequacy and abstraction for target applications • loosely coupled, horizontally scalable systems • Transparency – coordinated, but relatively independent systems – compatibility with previous platforms • application level distributed computing – scope and degree of location independence – peer-to-peer, application level protocols • Degree of Coupling – distributed middle-ware platforms – on how many things do distinct systems agree – how is that agreement achieved Advanced Architectures 3 Advanced Architectures 4 SMP systems and goals Symmetric Multi-Processors • Characterization: – multiple CPUs sharing memory and devices CPU 1 CPU 2 CPU 3 CPU 4 interrupt controller • Motivations: cache cache cache cache – price performance (lower price per MIP) shared memory & device busses – scalability (economical way to build huge systems) – perfect application transparency device device device controller controller controller • Example: memory – multi-core Intel CPUs – multi-socket mother boards Advanced Architectures 5 Advanced Architectures 6 1 6/7/2018 SMP Price/Performance SMP Operating System Design • a computer is much more than a CPU • one processor boots with power on – mother-board, disks, controllers, power supplies, case – it controls the starting of all other processors – CPU might cost 10-15% of the cost of the computer • same OS code runs in all processors • adding CPUs to a computer is very cost-effective – one physical copy in memory, shared by all CPUs – a second CPU yields cost of 1.1x, performance 1.9x • – a third CPU yields cost of 1.2x, performance 2.7x Each CPU has its own registers, cache, MMU – • same argument also applies at the chip level they must cooperatively share memory and devices – making a machine twice as fast is ever more difficult • ALL kernel operations must be Multi-Thread-Safe – adding more cores to the chip gets ever easier – protected by appropriate locks/semaphores • massive multi-processors are obvious direction – very fine grained locking to avoid contention Advanced Architectures 7 Advanced Architectures 8 SMP Parallelism The Challenge of SMP Performance • scheduling and load sharing • scalability depends on memory contention – each CPU can be running a different process – memory bandwidth is limited, can't handle all CPUs – just take the next ready process off the run-queue – most references satisfied from per-core cache – processes run in parallel – if too many requests go to memory, CPUs slow down – most processes don't interact (other than in kernel) • scalability depends on lock contention • serialization – waiting for spin-locks wastes time – mutual exclusion achieved by locks in shared memory – context switches waiting for kernel locks waste time – locks can be maintained with atomic instructions • contention wastes cycles, reduces throughput – spin locks acceptable for VERY short critical sections – 2 CPUs might deliver only 1.9x performance – if a process blocks, that CPU finds next ready process – 3 CPUs might deliver only 2.7x performance Advanced Architectures 9 Advanced Architectures 10 Managing Memory Contention Non-Uniform Memory Architecture • Fast n-way memory is very expensive Symmetric Multi-Processors – without it, memory contention taxes performance CPU n CPU n+1 – cost/complexity limits how many CPUs we can add local local cache memory cache memory • Non-Uniform Memory Architectures (NUMA) – each CPU has its own memory PCI bridge PCI bridge • each CPU has fast path to its own memory PCI bus PCI bus – connected by a Scalable Coherent Interconnect CC NUMA device device CC NUMA device device • a very fast , very local network between memories interface controller controller interface controller controller • accessing memory over the SCI may be 3-20x slower – these interconnects can be highly scalable Scalable Coherent Interconnect (e.g. Intel Quick Path Interconnect) Advanced Architectures 11 Advanced Architectures 12 2 6/7/2018 OS design for NUMA systems Single System Image (SSI) Clusters • it is all about local memory hit rates • Characterization: – every outside reference costs us 3-20x performance – a group of seemingly independent computers – we need 75-95% hit rate just to break even collaborating to provide SMP-like transparency • • How can the OS ensure high hit-rates? Motivation: – – replicate shared code pages in each CPU's memory higher reliability, availability than SMP/NUMA – – assign processes to CPUs, allocate all memory there more scalable than SMP/NUMA – – migrate processes to achieve load balancing excellent application transparency – spread kernel resources among all the CPUs • Examples: – attempt to preferentially allocate local resources – Locus, MicroSoft Wolf-Pack, OpenSSI – migrate resource ownership to CPU that is using it – Oracle Parallel Server Advanced Architectures 13 Advanced Architectures 14 Modern Clustered Architecture OS design for SSI clustering • all nodes agree on the state of all OS resources geographic fail over switch switch – file systems, processes, devices, locks IPC ports – any process can operate on any object, transparently ethernet request replication • they achieve this by exchanging messages SMP SMP SMP SMP system #1 system #2 system #3 system #4 – advising one-another of all changes to resources • each OS's internal state mirrors the global state dual ported synchronous FC replication dual ported optional – request execution of node-specific requests primary site RAID RAID back-up site • node-specific requests are forwarded to owning node Active systems service independent requests in parallel. They cooperate to maintain • implementation is large, complex, difficult shared global locks, and are prepared to take over partner’s work in case of failure. State replication to a back-up site is handled by external mechanisms. • the exchange of messages can be very expensive Advanced Architectures 15 Advanced Architectures 16 SSI Clustered Performance Lessons Learned • clever implementation can minimize overhead • consensus protocols are expensive – 10-20% overall is not uncommon, can be much worse – they converge slowly and scale poorly • • complete transparency systems have a great many resources – – even very complex applications "just work" resource change notifications are expensive • – they do not have to be made "network aware" location transparency encouraged non-locality – remote resource use is much more expensive • good robustness • a greatly complicated operating system – when one node fails, others notice and take-over – distributed objects are more complex to manage – often, applications won't even notice the failure – complex optimizations to reduce the added overheads • nice for application developers and customers – new modes of failure w/complex recovery procedures – but they are complex, and not particularly scalable • Bottom Line: Deutsch was right! Advanced Architectures 17 Advanced Architectures 18 3 6/7/2018 Loosely Coupled Systems Horizontal Scalability w/HA • Characterization: WAN to clients – a parallel group of independent computers load balancing switch – serving similar but independent requests w/fail-over – minimal coordination and cooperation required … web web web web web app app app app app … D1 • Motivation: server server server server server server server server server server – scalability and price performance – availability – if protocol permits stateless servers content HA – distribution database D2 ease of management, reconfigurable capacity server server • Examples: – web servers, Google search farm, Hadoop Advanced Architectures 19 Advanced Architectures 20 (elements of architecture) Horizontally scaled performance • farm of independent servers • individual servers are very inexpensive – servers run same software, serve different requests – blade servers may be only $100-$200 each – may share a common back-end database • scalability is excellent – • front-ending switch 100 servers deliver approximately 100x performance • – distributes incoming requests among available servers service availability is excellent – front-end automatically bypasses failed servers – can do both load balancing and fail-over – stateless servers and client retries fail-over easily • service protocol • the challenge is managing thousands of servers – stateless servers and idempotent operations – automated installation, global configuration services – successive requests may be sent to different servers – self monitoring, self-healing systems Advanced Architectures 21 Advanced Architectures 22 Limited Transparency Clusters Limited LocationTransparency
Load more

Recommended publications
  • Distributed & Cloud Computing: Lecture 1
    Distributed & cloud computing: Lecture 1 Chokchai “Box” Leangsuksun SWECO Endowed Professor, Computer Science Louisiana Tech University [email protected] CTO, PB Tech International Inc. [email protected] Class Info ■" Class Hours: noon-1:50 pm T-Th ! ■" Dr. Box’s & Contact Info: ●"www.latech.edu/~box ●"Phone 318-257-3291 ●"Email: [email protected] Class Info ■" Main Text: ! ●" 1) Distributed Systems: Concepts and Design By Coulouris, Dollimore, Kindberg and Blair Edition 5, © Addison-Wesley 2012 ●" 2) related publications & online literatures Class Info ■" Objective: ! ●"the theory, design and implementation of distributed systems ●"Discussion on abstractions, Concepts and current systems/applications ●"Best practice on Research on DS & cloud computing Course Contents ! •" The Characterization of distributed computing and cloud computing. •" System Models. •" Networking and inter-process communication. •" OS supports and Virtualization. •" RAS, Performance & Reliability Modeling. Security. !! Course Contents •" Introduction to Cloud Computing ! •" Various models and applications. •" Deployment models •" Service models (SaaS, PaaS, IaaS, Xaas) •" Public Cloud: Amazon AWS •" Private Cloud: openStack. •" Cost models ( between cloud vs host your own) ! •" ! !!! Course Contents case studies! ! •" Introduction to HPC! •" Multicore & openMP! •" Manycore, GPGPU & CUDA! •" Cloud-based EKG system! •" Distributed Object & Web Services (if time allows)! •" Distributed File system (if time allows)! •" Replication & Disaster Recovery Preparedness! !!!!!! Important Dates ! ■" Apr 10: Mid Term exam ■" Apr 22: term paper & presentation due ■" May 15: Final exam Evaluations ! ■" 20% Lab Exercises/Quizzes ■" 20% Programming Assignments ■" 10% term paper ■" 20% Midterm Exam ■" 30% Final Exam Intro to Distributed Computing •" Distributed System Definitions. •" Distributed Systems Examples: –" The Internet. –" Intranets. –" Mobile Computing –" Cloud Computing •" Resource Sharing. •" The World Wide Web. •" Distributed Systems Challenges. Based on Mostafa’s lecture 10 Distributed Systems 0.
    [Show full text]
  • Distributed Algorithms with Theoretic Scalability Analysis of Radial and Looped Load flows for Power Distribution Systems
    Electric Power Systems Research 65 (2003) 169Á/177 www.elsevier.com/locate/epsr Distributed algorithms with theoretic scalability analysis of radial and looped load flows for power distribution systems Fangxing Li *, Robert P. Broadwater ECE Department Virginia Tech, Blacksburg, VA 24060, USA Received 15 April 2002; received in revised form 14 December 2002; accepted 16 December 2002 Abstract This paper presents distributed algorithms for both radial and looped load flows for unbalanced, multi-phase power distribution systems. The distributed algorithms are developed from tree-based sequential algorithms. Formulas of scalability for the distributed algorithms are presented. It is shown that computation time dominates communication time in the distributed computing model. This provides benefits to real-time load flow calculations, network reconfigurations, and optimization studies that rely on load flow calculations. Also, test results match the predictions of derived formulas. This shows the formulas can be used to predict the computation time when additional processors are involved. # 2003 Elsevier Science B.V. All rights reserved. Keywords: Distributed computing; Scalability analysis; Radial load flow; Looped load flow; Power distribution systems 1. Introduction Also, the method presented in Ref. [10] was tested in radial distribution systems with no more than 528 buses. Parallel and distributed computing has been applied More recent works [11Á/14] presented distributed to many scientific and engineering computations such as implementations for power flows or power flow based weather forecasting and nuclear simulations [1,2]. It also algorithms like optimizations and contingency analysis. has been applied to power system analysis calculations These works also targeted power transmission systems. [3Á/14].
    [Show full text]
  • Grid Computing: What Is It, and Why Do I Care?*
    Grid Computing: What Is It, and Why Do I Care?* Ken MacInnis <[email protected]> * Or, “Mi caja es su caja!” (c) Ken MacInnis 2004 1 Outline Introduction and Motivation Examples Architecture, Components, Tools Lessons Learned and The Future Questions? (c) Ken MacInnis 2004 2 What is “grid computing”? Many different definitions: Utility computing Cycles for sale Distributed computing distributed.net RC5, SETI@Home High-performance resource sharing Clusters, storage, visualization, networking “We will probably see the spread of ‘computer utilities’, which, like present electric and telephone utilities, will service individual homes and offices across the country.” Len Kleinrock (1969) The word “grid” doesn’t equal Grid Computing: Sun Grid Engine is a mere scheduler! (c) Ken MacInnis 2004 3 Better definitions: Common protocols allowing large problems to be solved in a distributed multi-resource multi-user environment. “A computational grid is a hardware and software infrastructure that provides dependable, consistent, pervasive, and inexpensive access to high-end computational capabilities.” Kesselman & Foster (1998) “…coordinated resource sharing and problem solving in dynamic, multi- institutional virtual organizations.” Kesselman, Foster, Tuecke (2000) (c) Ken MacInnis 2004 4 New Challenges for Computing Grid computing evolved out of a need to share resources Flexible, ever-changing “virtual organizations” High-energy physics, astronomy, more Differing site policies with common needs Disparate computing needs
    [Show full text]
  • Computing at Massive Scale: Scalability and Dependability Challenges
    This is a repository copy of Computing at massive scale: Scalability and dependability challenges. White Rose Research Online URL for this paper: http://eprints.whiterose.ac.uk/105671/ Version: Accepted Version Proceedings Paper: Yang, R and Xu, J orcid.org/0000-0002-4598-167X (2016) Computing at massive scale: Scalability and dependability challenges. In: Proceedings - 2016 IEEE Symposium on Service-Oriented System Engineering, SOSE 2016. 2016 IEEE Symposium on Service-Oriented System Engineering (SOSE), 29 Mar - 02 Apr 2016, Oxford, United Kingdom. IEEE , pp. 386-397. ISBN 9781509022533 https://doi.org/10.1109/SOSE.2016.73 (c) 2016 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other users, including reprinting/ republishing this material for advertising or promotional purposes, creating new collective works for resale or redistribution to servers or lists, or reuse of any copyrighted components of this work in other works. Reuse Unless indicated otherwise, fulltext items are protected by copyright with all rights reserved. The copyright exception in section 29 of the Copyright, Designs and Patents Act 1988 allows the making of a single copy solely for the purpose of non-commercial research or private study within the limits of fair dealing. The publisher or other rights-holder may allow further reproduction and re-use of this version - refer to the White Rose Research Online record for this item. Where records identify the publisher as the copyright holder, users can verify any specific terms of use on the publisher’s website. Takedown If you consider content in White Rose Research Online to be in breach of UK law, please notify us by emailing [email protected] including the URL of the record and the reason for the withdrawal request.
    [Show full text]
  • Lecture 18: Parallel and Distributed Systems
    Lecture 18: Parallel and What is a distributed system? Distributed Systems n From various textbooks: u “A distributed system is a collection of independent computers n Classification of parallel/distributed that appear to the users of the system as a single computer.” architectures u “A distributed system consists of a collection of autonomous n SMPs computers linked to a computer network and equipped with n Distributed systems distributed system software.” u n Clusters “A distributed system is a collection of processors that do not share memory or a clock.” u “Distributed systems is a term used to define a wide range of computer systems from a weakly-coupled system such as wide area networks, to very strongly coupled systems such as multiprocessor systems.” 1 2 What is a distributed system?(cont.) What is a distributed system?(cont.) n Michael Flynn (1966) P1 P2 P3 u n A distributed system is Network SISD — single instruction, single data u SIMD — single instruction, multiple data a set of physically separate u processors connected by MISD — multiple instruction, single data u MIMD — multiple instruction, multiple data one or more P4 P5 communication links n More recent (Stallings, 1993) n Is every system with >2 computers a distributed system?? u Email, ftp, telnet, world-wide-web u Network printer access, network file access, network file backup u We don’t usually consider these to be distributed systems… 3 4 MIMD Classification parallel and distributed Classification of the OS of MIMD computers Architectures tightly loosely coupled
    [Show full text]
  • Hybrid Computing – Co-Processors/Accelerators Power-Aware Computing – Performance of Applications Kernels
    C-DAC Four Days Technology Workshop ON Hybrid Computing – Co-Processors/Accelerators Power-aware Computing – Performance of Applications Kernels hyPACK-2013 (Mode-1:Multi-Core) Lecture Topic : Multi-Core Processors : Algorithms & Applications Overview - Part-II Venue : CMSD, UoHYD ; Date : October 15-18, 2013 C-DAC hyPACK-2013 Multi-Core Processors : Alg.& Appls Overview Part-II 1 Killer Apps of Tomorrow Application-Driven Architectures: Analyzing Tera-Scale Workloads Workload convergence The basic algorithms shared by these high-end workloads Platform implications How workload analysis guides future architectures Programmer productivity Optimized architectures will ease the development of software Call to Action Benchmark suites in critical need for redress C-DAC hyPACK-2013 Multi-Core Processors : Alg.& Appls Overview Part-II 2 Emerging “Killer Apps of Tomorrow” Parallel Algorithms Design Static and Load Balancing Mapping for load balancing Minimizing Interaction Overheads in parallel algorithms design Data Sharing Overheads C-DAC hyPACK-2013 Multi-Core Processors : Alg.& Appls Overview Part-II 3 Emerging “Killer Apps of Tomorrow” Parallel Algorithms Design (Contd…) General Techniques for Choosing Right Parallel Algorithm Maximize data locality Minimize volume of data Minimize frequency of Interactions Overlapping computations with interactions. Data replication Minimize construction and Hot spots. Use highly optimized collective interaction operations. Collective data transfers and computations • Maximize Concurrency.
    [Show full text]
  • Cost Model: Work, Span and Parallelism 1 Overview of Cilk
    CSE 539 01/15/2015 Cost Model: Work, Span and Parallelism Lecture 2 Scribe: Angelina Lee Outline of this lecture: 1. Overview of Cilk 2. The dag computation model 3. Performance measures 4. A simple greedy scheduler 1 Overview of Cilk A brief history of Cilk technology In this class, we will overview various parallel programming technology developed. One of the technology we will examine closely is Cilk, a C/C++ based concurrency platform. Before we overview the development and implementation of Cilk, we shall first overview a brief history of Cilk technology to account for where the major concepts originate. Cilk technology has developed and evolved over more than 15 years since its origin at MIT. The text under this subheading is partially abstracted from the “Cilk” entry in Encyclopedia of Distributed Computing [14] with the author’s consent. I invite interested readers to go through the original entry for a more complete review of the history. Cilk (pronounced “silk”) is a linguistic and runtime technology for algorithmic multithreaded programming originally developed at MIT. The philosophy behind Cilk is that a programmer should concentrate on structuring her or his program to expose parallelism and exploit locality, leaving Cilk’s runtime system with the responsibility of scheduling the computation to run effi- ciently on a given platform. The Cilk runtime system takes care of details like load balancing, synchronization, and communication protocols. Cilk is algorithmic in that the runtime system guarantees efficient and predictable performance. Important milestones in Cilk technology include the original Cilk-1 [2,1, 11], 1 Cilk-5 [7, 18,5, 17], and the commercial Cilk++ [15] by Cilk Arts.
    [Show full text]
  • Difference Between Grid Computing Vs. Distributed Computing
    Difference between Grid Computing Vs. Distributed Computing Definition of Distributed Computing Distributed Computing is an environment in which a group of independent and geographically dispersed computer systems take part to solve a complex problem, each by solving a part of solution and then combining the result from all computers. These systems are loosely coupled systems coordinately working for a common goal. It can be defined as 1. A computing system in which services are provided by a pool of computers collaborating over a network. 2. A computing environment that may involve computers of differing architectures and data representation formats that share data and system resources. Distributed Computing, or the use of a computational cluster, is defined as the application of resources from multiple computers, networked in a single environment, to a single problem at the same time - usually to a scientific or technical problem that requires a great number of computer processing cycles or access to large amounts of data. Picture: The concept of distributed computing is simple -- pull together and employ all available resources to speed up computing. The key distinction between distributed computing and grid computing is mainly the way resources are managed. Distributed computing uses a centralized resource manager and all nodes cooperatively work together as a single unified resource or a system. Grid computingutilizes a structure where each node has its own resource manager and the system does not act as a single unit. Definition of Grid Computing The Basic idea between Grid Computing is to utilize the ideal CPU cycles and storage of millions of computer systems across a worldwide network function as a flexible, pervasive, and inexpensive accessible pool that could be harnessed by anyone who needs it, similar to the way power companies and their users share the electrical grid.
    [Show full text]
  • Distributed GPU-Based K-Means Algorithm for Data-Intensive Applications: Large-Sized Image Segmentation Case
    (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 8, No. 12, 2017 Distributed GPU-Based K-Means Algorithm for Data-Intensive Applications: Large-Sized Image Segmentation Case Hicham Fakhi*, Omar Bouattane, Mohamed Youssfi, Hassan Ouajji Signals, Distributed Systems and Artificial Intelligence Laboratory ENSET, University Hassan II, Casablanca, Morocco Abstract—K-means is a compute-intensive iterative Nonetheless, there are two important factors to consider algorithm. Its use in a complex scenario is cumbersome, when doing image segmentation. First is the number of images specifically in data-intensive applications. In order to accelerate to be processed in a given use case. Second is the image quality the K-means running time for data-intensive application, such as which has known an important evolution during the last few large sized image segmentation, we use a distributed multi-agent years, i.e., the number of pixels that make up an image has system accelerated by GPUs. In this K-means version, the input been multiplied by 200 from the 720 x 480 pixels to 9600 x image data are divided into subsets of image data which can be 7200 pixels. This has resulted in a much better definition of the performed independently on GPUs. In each GPU, we offloaded image (more detail visibility) and more nuances in the colors the data assignment and the K-centroids recalculation steps of and shades. the K-means algorithm for a massively parallel processing. We have implemented this K-means version on the Nvidia GPU with Thus, during last decade, image processing techniques have Compute Unified Device Architecture.
    [Show full text]
  • Parallel Sorting Algorithms in Intel Cilk Plus
    International Journal of Hybrid Information Technology Vol.7, No.2 (2014), pp.151-164 http://dx.doi.org/10.14257/ijhit.2014.7.2.15 Multi-Core Program Optimization: Parallel Sorting Algorithms in Intel Cilk Plus Sabahat Saleem1, M. IkramUllah Lali1, M. Saqib Nawaz1* and Abou Bakar Nauman2 1Department of Computer Science & IT, University of Sargodha, Sargodha, Pakistan 2Department of CS & IT, Sarhad University of Science & IT, Peshawar, Pakistan * [email protected] Abstract New performance leaps has been achieved with multiprogramming and multi-core systems. Present parallel programming techniques and environment needs significant changes in programs to accomplish parallelism and also constitute complex, confusing and error-prone constructs and rules. Intel Cilk Plus is a C based computing system that presents a straight forward and well-structured model for the development, verification and analysis of multi- core and parallel programming. In this article, two programs are developed using Intel Cilk Plus. Two sequential sorting programs in C/C++ languageOnly. are ILLEGAL. converted to multi-core programs in Intel Cilk Plus framework to achieve parallelism and better performance. Converted program in Cilk Plus is then checked for various conditionsis using tools of Cilk and after that, comparison of performance and speedup achieved over the single-core sequential program is discussed and reported. Keywords: Single-Core, Multi-Core, Cilk, Sortingfile Algorithms, Quick Sort, Merge Sort, Parallelism, Speedup Version 1. Introduction this With current technology, improving the number of transistors does not increase the computing capability of a system.by One of the new techniques used to increase the performance of a system is the development of multi-core processors.
    [Show full text]
  • Operating System Architecture Based on Distributed Objects
    located in different nodes. However, A. Yu. Burtsev, L. B. Ryzhyk, Yu. O. efficient access to resources can oc- Timoshenko cur only locally. E1 enables access locality by making an object physi- OPERATING SYSTEM ARCHITECTURE BASED ON cally distributed, i.e. storing a com- DISTRIBUTED OBJECTS plete or partial copy of its state in several network nodes. The internal organization of an object allows mini- Introduction mizing interactions with remote nodes during execution of operations on ob- Today the computing power of the jects, thus neutralizing the influence networks of workstations has become of communication latencies on overall comparable to that of modern supercom- system performance. puters. The practical implementation Load balancing. The E1 operating of high-performance distributed compu- system supports load balancing, i.e. tations obviously requires a special dynamic distribution of computational software platform. At present, this payload between system nodes. platform is usually implemented by ap- Support for redundancy mechanisms. plication-level parallel programming E1 allows applications to transpar- libraries, providing software develop- ently use distributed algorithms for ers with distributed communication and redundant data storage and execution. synchronization facilities [1,2]. Our Component model support. E1 has a claim is that the problem can be ad- component architecture and supports dressed in full measure by an operat- component-oriented software develop- ing system providing processes with ment paradigm. Both operating system transparent access to all resources of services and application software are the distributed system. Such operating developed within the framework of a system would allow users and applica- single E1 component model. The compo- tion programmers to think of the com- nent model provides the following ser- puter network as a single multiproces- vices: sor with common memory.
    [Show full text]
  • A Distributed Multi-GPU System for Fast Graph Processing
    A Distributed Multi-GPU System for Fast Graph Processing Zhihao Jia Yongkee Kwon Galen Shipman Stanford University UT Austin LANL [email protected] [email protected] [email protected] Pat McCormick Mattan Erez Alex Aiken LANL UT Austin Stanford University [email protected] [email protected] [email protected] ABSTRACT Network We present Lux, a distributed multi-GPU system that achieves CPUs CPUs DRAM DRAM fast graph processing by exploiting the aggregate memory Zero-Copy Memory Zero-Copy Memory bandwidth of multiple GPUs and taking advantage of lo- cality in the memory hierarchy of multi-GPU clusters. Lux PCI-e Switch / NVLink PCI-e Switch / NVLink provides two execution models that optimize algorithmic ef- Device Memory Device Memory Device Memory Device Memory ficiency and enable important GPU optimizations, respec- Shared Shared Shared Shared Memory Memory Memory Memory tively. Lux also uses a novel dynamic load balancing strat- GPU1 GPUN GPU1 GPUN egy that is cheap and achieves good load balance across Compute Node Compute Node GPUs. In addition, we present a performance model that Figure 1: Multi-GPU node architecture. quantitatively predicts the execution times and automati- cally selects the runtime configurations for Lux applications. 104 DRAM GPU Device Memory Experiments show that Lux achieves up to 20× speedup over GPU Shared Memory 103 GPU Accessing Zero-Copy Memory 432.1 435.0 state-of-the-art shared memory systems and up to two or- 354.4 356.5 ders of magnitude speedup over distributed systems. 150.0 163.0 102 PVLDB Reference Format: 15.0 c 101 8.7 Zhihao Jia, Yongkee Kwon, Galen Shipman, Pat M Cormick, Bandwidth (GB/s) 3.0 3.0 1.5 Mattan Erez, and Alex Aiken.
    [Show full text]