DOCSLIB.ORG

HOP-MAP: Efficient Message Passing with High Order Potentials

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
HOP-MAP: Efficient Message Passing with High Order Potentials HOP-MAP: Efficient Message Passing with High Order Potentials Daniel Tarlow Inmar E. Givoni Richard S. Zemel Dept. of Computer Science Probabilistic & Statistical Dept. of Computer Science University of Toronto Inference Group University of Toronto University of Toronto Abstract When potentials begin to range over a large subset of variables, however, these methods quickly break down: There is a growing interest in building prob- for the general problem, message updates from a high abilistic models with high order potentials order clique take time exponential in the size of the (HOPs), or interactions, among discrete vari- clique. One approach in such cases is to transform ables. Message passing inference in such the problem into a pairwise problem by adding aux- models generally takes time exponential in iliary variables. In the worst case, this will increase the size of the interaction, but in some cases the problem size exponentially. Rother et al. (2009) maximum a posteriori (MAP) inference can give transformations that are practical for the spe- be carried out efficiently. We build upon such cial case of sparse potentials. Alternatively, special- results, introducing two new classes, includ- purpose computations can sometimes be performed ing composite HOPs that allow us to flexibly directly on the original high order|typically factor combine tractable HOPs using simple logi- graph|representation (Givoni & Frey, 2009), which cal switching rules. We present efficient mes- is the approach we take in this paper. This strategy sage update algorithms for the new HOPs, applies beyond sparse potentials and is applicable for and we improve upon the efficiency of mes- a wide range of special potential structures. Unfortu- sage updates for a general class of existing nately, for a given potential it is typically not imme- HOPs. Importantly, we present both new diately clear whether it has tractable structure. Even and existing HOPs in a common represen- if it does, message updates are case-specific and must tation; performing inference with any combi- be specially derived. nation of these HOPs requires no change of Our goal is to be able to use a broad range of high representations or new derivations. order potentials (HOPs) generically within MAP mes- sage passing algorithms as easily as we use low order tabular potentials. We see three issues that must be 1 INTRODUCTION addressed: Probabilistic graphical models are powerful tools due 1. Message updates need to be computed efficiently to their representational power, and also due to general even when factors contain very large scopes. purpose algorithms that can be applied to any (low or- der) graphical model. For a broad range of problems, 2. It should be easy to recognize when problems con- we can formulate a model in terms of graph struc- tain tractable high order structure. tures and standard potentials. Then, without further derivation, we can automatically perform inference in 3. HOP constructions should be flexible and re- the model. In particular, when the aim is to find a suable, not requiring new problem-specific deriva- most likely configuration of variables (MAP), a range tions and implementations on each use. of efficient message passing algorithms can be applied (Wainwright et al., 2005; Werner, 2008; Globerson & In Section 3, we describe the classes of atomic, building Jaakkola, 2008). block HOPs that we consider, cardinality-based and order-based. Cardinality-based potentials have been Appearing in Proceedings of the 13th International Con- used in several related works, and efficient message ference on Artificial Intelligence and Statistics (AISTATS) computations exist for both the general case and sev- 2010, Chia Laguna Resort, Sardinia, Italy. Volume 9 of eral restricted classes. Our first contribution, however, JMLR: W&CP 9. Copyright 2010 by the authors. is showing that the efficiency can be improved even 812 HOP-MAP: Efficient Message Passing with High Order Potentials beyond existing efficient computations (Potetz, 2007; sent HOPs over subsets. θj : hj 2 f0; 1g ! R and jcj Tarlow et al., 2008). Our algorithm computes all mes- θc : hc 2 f0; 1g ! R assign a real value to a variable N sages in O(N log N) time, a factor of log N better than or subset assignment, respectively. The potentials we existing approaches. For the novel class of order-based present can range over any number of variables, so we potentials, we present equally efficient algorithms. For use N to generically represent jcj. the atomic potentials we consider, finding the optimal A factor graph, then, defines a (log) likelihood that assignment takes O(N) or O(N log N) time. We show takes the form that our algorithms can compute all N outgoing mes- n sages in the same asymptotic time. X X L(h) = θj(hj) + θc(hc) (1) Next, analogous in spirit to disciplined convex pro- j=1 c2C gramming, which allows many of the manipulations and transformations required to analyze and solve con- The MAP inference problem|to find a setting vex programs to be automated (Grant et al., 2006), we of h that maximizes the likelihood hOP T = introduce two types of composite HOPs, which allow us arg maxh L(h)|is NP-hard for general loopy graphs, to build more complex HOPs by applying composition and we typically must resort to approximate optimiza- rules based on maximization or simple logical switches. tion methods. A complex HOP can be recognized to be tractable by decomposing it into atomic tractable units combined 2.1 MAP MESSAGE PASSING with allowed composition rules. Importantly, once ex- pressed as a composition, the message updates of com- Max-product belief propagation (MPBP) is an itera- posite HOPs can automatically be computed from the tive, local, message passing algorithm that can be used message computations of the atomic HOPs. to find the MAP configuration of a probability distri- bution specified by a tree-structured graphical model. Our final, more subtle contribution is the particular binary representation, message normalization scheme, When working in log space, the algorithm is known as and caching strategy for computing all outgoing mes- max-sum, and the updates involve sending messages sages from a factor at once. This \framework," which from factors to variables, we refer to throughout, is not novel, but it is also not 2 3 X the standard. Each part has a purpose: the binary rep- 0 m~ θc!hj (hj) = max 4θc(hc) + m~ hj0 !θc (hj )5 , hcnfhj g resentation exposes more structure in potentials; the j02cjj06=j message normalization scheme yields simpler deriva- tions; and the caching strategy leads to more efficient and from variables to factors,m ~ h !θ (hj) = P j c 0 m~ (h ). After a forward and back- algorithms. It should be straightforward to apply this c 2N (hj )nc θc0 !hj j strategy to other structured potentials to build new ward pass sending messages to and from a root atomic building block HOPs. node, optimal assignments can be decoded from OP T beliefs, hj = arg maxhj b(hj), where b(hj) = Section 4 shows that well-known and novel graph con- P 0 m~ θ 0 !h (hj). In tree-structured graphs, be- structions are easily expressible using the vocabulary c 2N (hj ) c j liefs defined in this way give (up to a constant) max- of potentials discussed, and that once a model is ex- marginals:Φ = max L(h). In loopy graphs, pressed in this framework, it can be used in a variety j;a hjhj =a beliefs produce pseudo max-marginals, which do not of MAP inference procedures. In Section 5 we present account for the loopy graph structure but can be de- experimental results that illustrate the ease of con- coded to give approximate solutions, which have been structing models of interest with this formulation. shown to be useful in practice. 2 REPRESENTATION 2.2 BINARY REPRESENTATION Note that we have restricted our attention to binary We work with a factor graph representation, which is variable problems. To represent multinomial variables, a bipartite graph consisting of variable nodes, h = we apply a simple transformation, converting variables fh ; : : : ; h g, and factor nodes. Let N (h ) be the 1 n j with L states to binary variables with a 1-of-L con- neighbors of variable h in the graph. Factors, or j straint ensuring that exactly one variable in the set is potentials, θ = fθ ; : : : ; θ ; θ ; : : : ; θ g, define in- 1 n n+1 n+k on. teractions over individual variables hj and subsets of variables C = fc1; : : : ; ckg; c ⊆ fh1; : : : ; hng, which Since all variables are binary, we normalize messages are exactly the factor's neighbors in the factor graph. so that the entry for hj = 0 is always 0. To enforce this With slight abuse of notation, we use θj to represent constraint, we subtractm ~ θ !h (0) from both coordi- c j node potentials over single variables, and θc to repre- nates, giving us 0; m~ θc!hj (1) − m~ θc!hj (0) . We can 813 Daniel Tarlow, Inmar E. Givoni, Richard S. Zemel then drop the argument for mθc!hj (1) and use mθc!hj potentials) over cardinality in O(N log N) time by to represent the scalar message difference. Similarly, a simple sorting then greedy procedure. We could we assume θj(0) = 0 for all node potentials by setting compute max-marginals for the star graph naively in ~ ~ 2 θj = θj(1) = θj(1) − θj(0). O(N log N) time by running this procedure itera- tively fixing at each iteration one value of one variable. We are then working with message differences, where a positive value indicates a variable's preference to If we use MPBP, Potetz (2007) shows that a sin- be on, and a negative value indicates a variable's gle message can be approximately computed in O(N) preference to be off.
Load more

Recommended publications
  • Efficient Inter-Core Communications on Manycore Machines
    ZIMP: Efficient Inter-core Communications on Manycore Machines Pierre-Louis Aublin Sonia Ben Mokhtar Gilles Muller Vivien Quema´ Grenoble University CNRS - LIRIS INRIA CNRS - LIG Abstract—Modern computers have an increasing num- nication mechanisms enabling both one-to-one and one- ber of cores and, as exemplified by the recent Barrelfish to-many communications. Current operating systems operating system, the software they execute increasingly provide various inter-core communication mechanisms. resembles distributed, message-passing systems. To sup- port this evolution, there is a need for very efficient However it is not clear yet how these mechanisms be- inter-core communication mechanisms. Current operat- have on manycore processors, especially when messages ing systems provide various inter-core communication are intended to many recipient cores. mechanisms, but it is not clear yet how they behave In this report, we study seven communication mech- on manycore processors. In this report, we study seven anisms, which are considered state-of-the-art. More mechanisms, that are considered state-of-the-art. We show that these mechanisms have two main drawbacks that precisely, we study TCP, UDP and Unix domain sock- limit their efficiency: they perform costly memory copy ets, pipes, IPC and POSIX message queues. These operations and they do not provide efficient support mechanisms are supported by traditional operating sys- for one-to-many communications. We do thus propose tems such as Linux. We also study Barrelfish message ZIMP, a new inter-core communication mechanism that passing, the inter-process communication mechanism of implements zero-copy inter-core message communications and that efficiently handles one-to-many communications.
    [Show full text]
  • D-Bus, the Message Bus System Training Material
    Maemo Diablo D-Bus, The Message Bus System Training Material February 9, 2009 Contents 1 D-Bus, The Message Bus System 2 1.1 Introduction to D-Bus ......................... 2 1.2 D-Bus architecture and terminology ................ 3 1.3 Addressing and names in D-Bus .................. 4 1.4 Role of D-Bus in maemo ....................... 6 1.5 Programming directly with libdbus ................. 9 1 Chapter 1 D-Bus, The Message Bus System 1.1 Introduction to D-Bus D-Bus (the D originally stood for "Desktop") is a relatively new inter process communication (IPC) mechanism designed to be used as a unified middleware layer in free desktop environments. Some example projects where D-Bus is used are GNOME and Hildon. Compared to other middleware layers for IPC, D-Bus lacks many of the more refined (and complicated) features and for that reason, is faster and simpler. D-Bus does not directly compete with low level IPC mechanisms like sock- ets, shared memory or message queues. Each of these mechanisms have their uses, which normally do not overlap the ones in D-Bus. Instead, D-Bus aims to provide higher level functionality, like: Structured name spaces • Architecture independent data formatting • Support for the most common data elements in messages • A generic remote call interface with support for exceptions (errors) • A generic signalling interface to support "broadcast" type communication • Clear separation of per-user and system-wide scopes, which is important • when dealing with multi-user systems Not bound to any specific programming language (while providing a • design that readily maps to most higher level languages, via language specific bindings) The design of D-Bus benefits from the long experience of using other mid- dleware IPC solutions in the desktop arena and this has allowed the design to be optimised.
    [Show full text]
  • Message Passing Model
    Message Passing Model Giuseppe Anastasi [email protected] Pervasive Computing & Networking Lab. (PerLab) Dept. of Information Engineering, University of Pisa PerLab Based on original slides by Silberschatz, Galvin and Gagne Overview PerLab Message Passing Model Addressing Synchronization Example of IPC systems Message Passing Model 2 Operating Systems Objectives PerLab To introduce an alternative solution (to shared memory) for process cooperation To show pros and cons of message passing vs. shared memory To show some examples of message-based communication systems Message Passing Model 3 Operating Systems Inter-Process Communication (IPC) PerLab Message system – processes communicate with each other without resorting to shared variables. IPC facility provides two operations: send (message ) – fixed or variable message size receive (message ) If P and Q wish to communicate, they need to: establish a communication link between them exchange messages via send/receive The communication link is provided by the OS Message Passing Model 4 Operating Systems Implementation Issues PerLab Physical implementation Single-processor system Shared memory Multi-processor systems Hardware bus Distributed systems Networking System + Communication networks Message Passing Model 5 Operating Systems Implementation Issues PerLab Logical properties Can a link be associated with more than two processes? How many links can there be between every pair of communicating processes? What is the capacity of a link? Is the size of a message that the link can accommodate fixed or variable? Is a link unidirectional or bi-directional? Message Passing Model 6 Operating Systems Implementation Issues PerLab Other Aspects Addressing Synchronization Buffering Message Passing Model 7 Operating Systems Overview PerLab Message Passing Model Addressing Synchronization Example of IPC systems Message Passing Model 8 Operating Systems Direct Addressing PerLab Processes must name each other explicitly.
    [Show full text]
  • Message Passing
    COS 318: Operating Systems Message Passing Jaswinder Pal Singh Computer Science Department Princeton University (http://www.cs.princeton.edu/courses/cos318/) Sending A Message Within A Computer Across A Network P1 P2 Send() Recv() Send() Recv() Network OS Kernel OS OS COS461 2 Synchronous Message Passing (Within A System) Synchronous send: u Call send system call with M u send system call: l No buffer in kernel: block send( M ) recv( M ) l Copy M to kernel buffer Synchronous recv: u Call recv system call M u recv system call: l No M in kernel: block l Copy to user buffer How to manage kernel buffer? 3 API Issues u Message S l Buffer (addr) and size l Message type, buffer and size u Destination or source send(dest, msg) R l Direct address: node Id, process Id l Indirect address: mailbox, socket, recv(src, msg) channel, … 4 Direct Addressing Example Producer(){ Consumer(){ ... ... while (1) { for (i=0; i<N; i++) produce item; send(Producer, credit); recv(Consumer, &credit); while (1) { send(Consumer, item); recv(Producer, &item); } send(Producer, credit); } consume item; } } u Does this work? u Would it work with multiple producers and 1 consumer? u Would it work with 1 producer and multiple consumers? u What about multiple producers and multiple consumers? 5 Indirect Addressing Example Producer(){ Consumer(){ ... ... while (1) { for (i=0; i<N; i++) produce item; send(prodMbox, credit); recv(prodMbox, &credit); while (1) { send(consMbox, item); recv(consMbox, &item); } send(prodMbox, credit); } consume item; } } u Would it work with multiple producers and 1 consumer? u Would it work with 1 producer and multiple consumers? u What about multiple producers and multiple consumers? 6 Indirect Communication u Names l mailbox, socket, channel, … u Properties l Some allow one-to-one mbox (e.g.
    [Show full text]
  • Multiprocess Communication and Control Software for Humanoid Robots Neil T
    IEEE Robotics and Automation Magazine Multiprocess Communication and Control Software for Humanoid Robots Neil T. Dantam∗ Daniel M. Lofaroy Ayonga Hereidx Paul Y. Ohz Aaron D. Amesx Mike Stilman∗ I. Introduction orrect real-time software is vital for robots in safety-critical roles such as service and disaster response. These systems depend on software for Clocomotion, navigation, manipulation, and even seemingly innocuous tasks such as safely regulating battery voltage. A multi-process software design increases robustness by isolating errors to a single process, allowing the rest of the system to continue operating. This approach also assists with modularity and concurrency. For real-time tasks such as dynamic balance and force control of manipulators, it is critical to communicate the latest data sample with minimum latency. There are many communication approaches intended for both general purpose and real-time needs [19], [17], [13], [9], [15]. Typical methods focus on reliable communication or network-transparency and accept a trade-off of increased mes- sage latency or the potential to discard newer data. By focusing instead on the specific case of real-time communication on a single host, we reduce communication latency and guarantee access to the latest sample. We present a new Interprocess Communication (IPC) library, Ach,1 which addresses this need, and discuss its application for real-time, multiprocess control on three humanoid robots (Fig. 1). There are several design decisions that influenced this robot software and motivated development of the Ach library. First, to utilize decades of prior development and engineering, we implement our real-time system on top of a POSIX-like Operating System (OS)2.
    [Show full text]
  • Message Passing
    Message passing CS252 • Sending of messages under control of programmer – User-level/system level? Graduate Computer Architecture – Bulk transfers? Lecture 24 • How efficient is it to send and receive messages? – Speed of memory bus? First-level cache? Network Interface Design • Communication Model: Memory Consistency Models – Synchronous » Send completes after matching recv and source data sent » Receive completes after data transfer complete from matching send – Asynchronous Prof John D. Kubiatowicz » Send completes after send buffer may be reused http://www.cs.berkeley.edu/~kubitron/cs252 4/27/2009 cs252-S09, Lecture 24 2 Synchronous Message Passing Asynch. Message Passing: Optimistic Source Destination Source Destination (1) Initiate send Recv Psrc, local VA, len (1) Initiate send (2) Address translation on Psrc Send Pdest, local VA, len (2) Address translation (3) Local/remote check Send (P , local VA, len) (3) Local/remote check dest (4) Send-ready request Send-rdy req (4) Send data (5) Remote check for (5) Remote check for posted receive; on fail, Tag match allocate data buffer posted receive Wait Tag check Data-xfer req (assume success) Processor Allocate buffer Action? (6) Reply transaction Recv-rdy reply (7) Bulk data transfer Source VA Dest VA or ID Recv P , local VA, len Time src Data-xfer req Time • More powerful programming model • Constrained programming model. • Wildcard receive => non-deterministic • Deterministic! What happens when threads added? • Destination contention very limited. • Storage required within msg layer? • User/System boundary? 4/27/2009 cs252-S09, Lecture 24 3 4/27/2009 cs252-S09, Lecture 24 4 Asynch. Msg Passing: Conservative Features of Msg Passing Abstraction Destination • Source knows send data address, dest.
    [Show full text]
  • Inter-Process Communication
    QNX Inter-Process Communication QNX IPC 2008/01/08 R08 © 2006 QNX Software Systems. QNX and Momentics are registered trademarks of QNX Software Systems in certain jurisdictions NOTES: QNX, Momentics, Neutrino, Photon microGUI, and “Build a more reliable world” are registered trademarks in certain jurisdictions, and PhAB, Phindows, and Qnet are trademarks of QNX Software Systems Ltd. All other trademarks and trade names belong to their respective owners. 1 IntroductionIntroduction IPC: • Inter-Process Communication • two processes exchange: – data – control – notification of event occurrence IPC Thread Thread Process A Process B QNX IPC 2008/01/08 R08 2 © 2006, QNX Software Systems. QNX and Momentics are registered trademarks of QNX Software Systems in certain jurisdictions NOTES: 2 IntroductionIntroduction QNX Neutrino supports a wide variety of IPC: – QNX Core (API is unique to QNX or low-level) • includes: – QNX Neutrino messaging – QNX Neutrino pulses – shared memory – QNX asynchronous messaging • the focus of this course module – POSIX/Unix (well known, portable API’s ) • includes: – signals – shared memory – pipes (requires pipe process) – POSIX message queues (requires mqueue or mq process) – TCP/IP sockets (requires io-net process) • the focus of the POSIX IPC course module QNX IPC 2008/01/08 R08 3 © 2006, QNX Software Systems. QNX and Momentics are registered trademarks of QNX Software Systems in certain jurisdictions NOTES: API: Application Programming Interface 3 InterprocessInterprocess Communication Communication Topics: Message Passing Designing a Message Passing System (1) Pulses How a Client Finds a Server Server Cleanup Multi-Part Messages Designing a Message Passing System (2) Client Information Structure Issues Related to Priorities Designing a Message Passing System (3) Event Delivery Asynchronous Message Passing QNET Shared Memory Conclusion QNX IPC 2008/01/08 R08 4 © 2006, QNX Software Systems.
    [Show full text]
  • 1 Interprocess Communication
    COMP 242 Class Notes Section 3: Interprocess Communication 1 Interprocess Communication Readings: Chapter 7, Comer. An operating system provides interprocess communication to allow processes to exchange informa- tion. Interprocess communication is useful for creating cooperating processes. For instance an ‘ls’ process and a ‘more’ process can cooperate to produce a paged listing of a directory. There are several mechanisms for interprocess communication. We discuss some of these below. 1.1 Shared Memory Processes that share memory can exchange data by writing and reading shared variables. As an example, consider two processes p and q that share some variable s. Then p can communicate information to q by writing new data in s, which q can then read. The above discussion raises an important question. How does q know when p writes new information into s? In some cases, q does not need to know. For instance, if it is a load balancing program that simply looks at the current load in p’s machine stored in s. When it does need to know, it could poll, but polling puts undue burden on the cpu. Another possibility is that it could get a software interrupt, which we discuss below. A familiar alternative is to use semaphores or conditions. Process q could block till p changes s and sends a signal that unblocks q. However, these solutions would not allow q to (automatically) block if s cannot hold all the data it wants to write. (The programmer could manually implement a bounded buffer using semaphores.) Moreover, conventional shared memory is accessed by processes on a single machine, so it cannot be used for communicating information among remote processes.
    [Show full text]
  • Shared Memory Introduction
    12 Shared Memory Introduction 12.1 Introduction Shared memory is the fastest form of IPC available. Once the memory is mapped into the address space of the processes that are sharing the memory region, no kernel involvement occurs in passing data between the processes. What is normally required, however, is some form of synchronization between the processes that are storing and fetching information to and from the shared memory region. In Part 3, we discussed various forms of synchronization: mutexes, condition variables, read–write locks, record locks, and semaphores. What we mean by ‘‘no kernel involvement’’ is that the processes do not execute any sys- tem calls into the kernel to pass the data. Obviously, the kernel must establish the mem- ory mappings that allow the processes to share the memory, and then manage this memory over time (handle page faults, and the like). Consider the normal steps involved in the client–server file copying program that we used as an example for the various types of message passing (Figure 4.1). • The server reads from the input file. The file data is read by the kernel into its memory and then copied from the kernel to the process. • The server writes this data in a message, using a pipe, FIFO, or message queue. These forms of IPC normally require the data to be copied from the process to the kernel. We use the qualifier normally because Posix message queues can be implemented using memory-mapped I/O (the mmap function that we describe in this chapter), as we showed in Section 5.8 and as we show in the solution to Exercise 12.2.
    [Show full text]
  • The Client/Server Architecture DES-0005 Revision 4
    PNNL-22587 The Client/Server Architecture DES-0005 Revision 4 Charlie Hubbard February 2012 The Client/Server Architecture Charlie Hubbard DES-0005 Revision 4 February 2012 DISCLAIMER 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 Battelle Memorial Institute, 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. Reference herein to any specific commercial product, process, or service by trade name, trademark, manufacturer, or otherwise does not necessarily constitute or imply its endorsement, recommendation, or favoring by the United States Government or any agency thereof, or Battelle Memorial Institute. The views and opinions of authors expressed herein do not necessarily state or reflect those of the United States Government or any agency thereof. PACIFIC NORTHWEST NATIONAL LABORATORY operated by BATTELLE for the UNITED STATES DEPARTMENT OF ENERGY under Contract DE-AC05-76RL01830 Printed in the United States of America Available to DOE and DOE contractors from the Office of Scientific and Technical Information, P.O. Box 62, Oak Ridge, TN 37831-0062 ph: (865) 576-8401 fax: (865) 576-5728 email: [email protected] Available to the public from the National Technical Information Service,
    [Show full text]
  • MODELS of DISTRIBUTED SYSTEMS Basic Elements
    Distributed Systems Fö 2/3- 1 Distributed Systems Fö 2/3- 2 MODELS OF DISTRIBUTED SYSTEMS Basic Elements • Resources in a distributed system are shared between users. They are normally encapsulated within one of the computers and can be accessed from other computers by communication. 1. Architectural Models • Each resource is managed by a program, the resource manager; it offers a communication 2. Interaction Models interface enabling the resource to be deceased by its users. 3. Failure Models • Resource managers can be in general modelled as processes. If the system is designed according to an object- oriented methodology, resources are encapsulated in objects. Petru Eles, IDA, LiTH Petru Eles, IDA, LiTH Distributed Systems Fö 2/3- 3 Distributed Systems Fö 2/3- 4 Architectural Models Client - Server Model How are responsibilities distributed between system components and how are these components placed? ☞ The system is structured as a set of processes, called servers, that offer services to the users, called • Client-server model clients. • Proxy server • Peer processes (peer to peer) • The client-server model is usually based on a simple request/reply protocol, implemented with send/receive primitives or using remote procedure calls (RPC) or remote method invocation (RMI): - the client sends a request message to the server asking for some service; - the server does the work and returns a result (e.g. the data requested) or an error code if the work could not be performed. Petru Eles, IDA, LiTH Petru Eles, IDA, LiTH Distributed Systems Fö 2/3- 5 Distributed Systems Fö 2/3- 6 Client - Server Model (cont’d) Proxy Server ☞ A proxy server provides copies (replications) of resources which are managed by other servers.
    [Show full text]
  • A Ugni-Based Asynchronous Message-Driven Runtime System for Cray Supercomputers with Gemini Interconnect
    A uGNI-based Asynchronous Message-driven Runtime System for Cray Supercomputers with Gemini Interconnect Yanhua Sun, Gengbin Zheng, Laxmikant V. Kale´ Terry R. Jones Ryan Olson Department of Computer Science Computer Science and Mathematics Division Cray Inc University of Illinois at Urbana-Champaign Oak Ridge National Lab [email protected] Urbana, Illinois Oak Ridge, Tennessee fsun51, gzheng, [email protected] [email protected] Abstract—Gemini, the network for the new Cray XE/XK user-level Generic Network Interface (uGNI) [?] is pro- systems, features low latency, high bandwidth and strong scal- vided as a low-level interface for Gemini, on top of which ability. Its hardware support for remote direct memory access the Cray MPI is implemented. Gemini-based interconnects enables efficient implementation of the global address space programming languages. Although the user Generic Network have been deployed in two petaflop supercomputers, Cielo Interface (uGNI) provides a low-level interface for Gemini with at Los Alamos National Lab and Hopper at Lawrence support to the message-passing programming model (MPI), it Berkeley National Lab. Jaguar, one of the most powerful remains challenging to port alternative programming models supercomputers, is being upgraded to a XK6 system with with scalable performance. Gemini interconnect. The newly announced Blue Waters CHARM++ is an object-oriented message-driven program- ming model. Its applications have been shown to scale up to supercomputer, which will deliver sustained performance of the full Jaguar Cray XT machine. In this paper, we present 1 petaflop, will also be equipped with a Gemini interconnect. an implementation of this programming model on uGNI for As Cray supercomputers with Gemini interconnects become the Cray XE/XK systems.
    [Show full text]