- •Contents
- •List of Figures
- •List of Tables
- •Acknowledgments
- •Introduction to MPI
- •Overview and Goals
- •Background of MPI-1.0
- •Background of MPI-1.1, MPI-1.2, and MPI-2.0
- •Background of MPI-1.3 and MPI-2.1
- •Background of MPI-2.2
- •Who Should Use This Standard?
- •What Platforms Are Targets For Implementation?
- •What Is Included In The Standard?
- •What Is Not Included In The Standard?
- •Organization of this Document
- •MPI Terms and Conventions
- •Document Notation
- •Naming Conventions
- •Semantic Terms
- •Data Types
- •Opaque Objects
- •Array Arguments
- •State
- •Named Constants
- •Choice
- •Addresses
- •Language Binding
- •Deprecated Names and Functions
- •Fortran Binding Issues
- •C Binding Issues
- •C++ Binding Issues
- •Functions and Macros
- •Processes
- •Error Handling
- •Implementation Issues
- •Independence of Basic Runtime Routines
- •Interaction with Signals
- •Examples
- •Point-to-Point Communication
- •Introduction
- •Blocking Send and Receive Operations
- •Blocking Send
- •Message Data
- •Message Envelope
- •Blocking Receive
- •Return Status
- •Passing MPI_STATUS_IGNORE for Status
- •Data Type Matching and Data Conversion
- •Type Matching Rules
- •Type MPI_CHARACTER
- •Data Conversion
- •Communication Modes
- •Semantics of Point-to-Point Communication
- •Buffer Allocation and Usage
- •Nonblocking Communication
- •Communication Request Objects
- •Communication Initiation
- •Communication Completion
- •Semantics of Nonblocking Communications
- •Multiple Completions
- •Non-destructive Test of status
- •Probe and Cancel
- •Persistent Communication Requests
- •Send-Receive
- •Null Processes
- •Datatypes
- •Derived Datatypes
- •Type Constructors with Explicit Addresses
- •Datatype Constructors
- •Subarray Datatype Constructor
- •Distributed Array Datatype Constructor
- •Address and Size Functions
- •Lower-Bound and Upper-Bound Markers
- •Extent and Bounds of Datatypes
- •True Extent of Datatypes
- •Commit and Free
- •Duplicating a Datatype
- •Use of General Datatypes in Communication
- •Correct Use of Addresses
- •Decoding a Datatype
- •Examples
- •Pack and Unpack
- •Canonical MPI_PACK and MPI_UNPACK
- •Collective Communication
- •Introduction and Overview
- •Communicator Argument
- •Applying Collective Operations to Intercommunicators
- •Barrier Synchronization
- •Broadcast
- •Example using MPI_BCAST
- •Gather
- •Examples using MPI_GATHER, MPI_GATHERV
- •Scatter
- •Examples using MPI_SCATTER, MPI_SCATTERV
- •Example using MPI_ALLGATHER
- •All-to-All Scatter/Gather
- •Global Reduction Operations
- •Reduce
- •Signed Characters and Reductions
- •MINLOC and MAXLOC
- •All-Reduce
- •Process-local reduction
- •Reduce-Scatter
- •MPI_REDUCE_SCATTER_BLOCK
- •MPI_REDUCE_SCATTER
- •Scan
- •Inclusive Scan
- •Exclusive Scan
- •Example using MPI_SCAN
- •Correctness
- •Introduction
- •Features Needed to Support Libraries
- •MPI's Support for Libraries
- •Basic Concepts
- •Groups
- •Contexts
- •Intra-Communicators
- •Group Management
- •Group Accessors
- •Group Constructors
- •Group Destructors
- •Communicator Management
- •Communicator Accessors
- •Communicator Constructors
- •Communicator Destructors
- •Motivating Examples
- •Current Practice #1
- •Current Practice #2
- •(Approximate) Current Practice #3
- •Example #4
- •Library Example #1
- •Library Example #2
- •Inter-Communication
- •Inter-communicator Accessors
- •Inter-communicator Operations
- •Inter-Communication Examples
- •Caching
- •Functionality
- •Communicators
- •Windows
- •Datatypes
- •Error Class for Invalid Keyval
- •Attributes Example
- •Naming Objects
- •Formalizing the Loosely Synchronous Model
- •Basic Statements
- •Models of Execution
- •Static communicator allocation
- •Dynamic communicator allocation
- •The General case
- •Process Topologies
- •Introduction
- •Virtual Topologies
- •Embedding in MPI
- •Overview of the Functions
- •Topology Constructors
- •Cartesian Constructor
- •Cartesian Convenience Function: MPI_DIMS_CREATE
- •General (Graph) Constructor
- •Distributed (Graph) Constructor
- •Topology Inquiry Functions
- •Cartesian Shift Coordinates
- •Partitioning of Cartesian structures
- •Low-Level Topology Functions
- •An Application Example
- •MPI Environmental Management
- •Implementation Information
- •Version Inquiries
- •Environmental Inquiries
- •Tag Values
- •Host Rank
- •IO Rank
- •Clock Synchronization
- •Memory Allocation
- •Error Handling
- •Error Handlers for Communicators
- •Error Handlers for Windows
- •Error Handlers for Files
- •Freeing Errorhandlers and Retrieving Error Strings
- •Error Codes and Classes
- •Error Classes, Error Codes, and Error Handlers
- •Timers and Synchronization
- •Startup
- •Allowing User Functions at Process Termination
- •Determining Whether MPI Has Finished
- •Portable MPI Process Startup
- •The Info Object
- •Process Creation and Management
- •Introduction
- •The Dynamic Process Model
- •Starting Processes
- •The Runtime Environment
- •Process Manager Interface
- •Processes in MPI
- •Starting Processes and Establishing Communication
- •Reserved Keys
- •Spawn Example
- •Manager-worker Example, Using MPI_COMM_SPAWN.
- •Establishing Communication
- •Names, Addresses, Ports, and All That
- •Server Routines
- •Client Routines
- •Name Publishing
- •Reserved Key Values
- •Client/Server Examples
- •Ocean/Atmosphere - Relies on Name Publishing
- •Simple Client-Server Example.
- •Other Functionality
- •Universe Size
- •Singleton MPI_INIT
- •MPI_APPNUM
- •Releasing Connections
- •Another Way to Establish MPI Communication
- •One-Sided Communications
- •Introduction
- •Initialization
- •Window Creation
- •Window Attributes
- •Communication Calls
- •Examples
- •Accumulate Functions
- •Synchronization Calls
- •Fence
- •General Active Target Synchronization
- •Lock
- •Assertions
- •Examples
- •Error Handling
- •Error Handlers
- •Error Classes
- •Semantics and Correctness
- •Atomicity
- •Progress
- •Registers and Compiler Optimizations
- •External Interfaces
- •Introduction
- •Generalized Requests
- •Examples
- •Associating Information with Status
- •MPI and Threads
- •General
- •Initialization
- •Introduction
- •File Manipulation
- •Opening a File
- •Closing a File
- •Deleting a File
- •Resizing a File
- •Preallocating Space for a File
- •Querying the Size of a File
- •Querying File Parameters
- •File Info
- •Reserved File Hints
- •File Views
- •Data Access
- •Data Access Routines
- •Positioning
- •Synchronism
- •Coordination
- •Data Access Conventions
- •Data Access with Individual File Pointers
- •Data Access with Shared File Pointers
- •Noncollective Operations
- •Collective Operations
- •Seek
- •Split Collective Data Access Routines
- •File Interoperability
- •Datatypes for File Interoperability
- •Extent Callback
- •Datarep Conversion Functions
- •Matching Data Representations
- •Consistency and Semantics
- •File Consistency
- •Random Access vs. Sequential Files
- •Progress
- •Collective File Operations
- •Type Matching
- •Logical vs. Physical File Layout
- •File Size
- •Examples
- •Asynchronous I/O
- •I/O Error Handling
- •I/O Error Classes
- •Examples
- •Subarray Filetype Constructor
- •Requirements
- •Discussion
- •Logic of the Design
- •Examples
- •MPI Library Implementation
- •Systems with Weak Symbols
- •Systems Without Weak Symbols
- •Complications
- •Multiple Counting
- •Linker Oddities
- •Multiple Levels of Interception
- •Deprecated Functions
- •Deprecated since MPI-2.0
- •Deprecated since MPI-2.2
- •Language Bindings
- •Overview
- •Design
- •C++ Classes for MPI
- •Class Member Functions for MPI
- •Semantics
- •C++ Datatypes
- •Communicators
- •Exceptions
- •Mixed-Language Operability
- •Problems With Fortran Bindings for MPI
- •Problems Due to Strong Typing
- •Problems Due to Data Copying and Sequence Association
- •Special Constants
- •Fortran 90 Derived Types
- •A Problem with Register Optimization
- •Basic Fortran Support
- •Extended Fortran Support
- •The mpi Module
- •No Type Mismatch Problems for Subroutines with Choice Arguments
- •Additional Support for Fortran Numeric Intrinsic Types
- •Language Interoperability
- •Introduction
- •Assumptions
- •Initialization
- •Transfer of Handles
- •Status
- •MPI Opaque Objects
- •Datatypes
- •Callback Functions
- •Error Handlers
- •Reduce Operations
- •Addresses
- •Attributes
- •Extra State
- •Constants
- •Interlanguage Communication
- •Language Bindings Summary
- •Groups, Contexts, Communicators, and Caching Fortran Bindings
- •External Interfaces C++ Bindings
- •Change-Log
- •Bibliography
- •Examples Index
- •MPI Declarations Index
- •MPI Function Index
1
2
22 |
CHAPTER 2. MPI TERMS AND CONVENTIONS |
2.7 Processes
3
4
5
An MPI program consists of autonomous processes, executing their own code, in an MIMD style. The codes executed by each process need not be identical. The processes communicate via calls to MPI communication primitives. Typically, each process executes in its own
6address space, although shared-memory implementations of MPI are possible.
7
8
9
This document speci es the behavior of a parallel program assuming that only MPI calls are used. The interaction of an MPI program with other possible means of communication, I/O, and process management is not speci ed. Unless otherwise stated in the
10speci cation of the standard, MPI places no requirements on the result of its interaction
11with external mechanisms that provide similar or equivalent functionality. This includes,
12but is not limited to, interactions with external mechanisms for process control, shared and
13remote memory access, le system access and control, interprocess communication, process
14signaling, and terminal I/O. High quality implementations should strive to make the results
15of such interactions intuitive to users, and attempt to document restrictions where deemed
16necessary.
17
18Advice to implementors. Implementations that support such additional mechanisms
19for functionality supported within MPI are expected to document how these interact
20with MPI. (End of advice to implementors.)
21
22
23
The interaction of MPI and threads is de ned in Section 12.4.
24 |
2.8 Error Handling |
|
|
25 |
|
26MPI provides the user with reliable message transmission. A message sent is always received
27correctly, and the user does not need to check for transmission errors, time-outs, or other
28error conditions. In other words, MPI does not provide mechanisms for dealing with failures
29in the communication system. If the MPI implementation is built on an unreliable underly-
30ing mechanism, then it is the job of the implementor of the MPI subsystem to insulate the
31user from this unreliability, or to re ect unrecoverable errors as failures. Whenever possible,
32such failures will be re ected as errors in the relevant communication call. Similarly, MPI
33itself provides no mechanisms for handling processor failures.
34Of course, MPI programs may still be erroneous. A program error can occur when
35an MPI call is made with an incorrect argument (non-existing destination in a send oper-
36ation, bu er too small in a receive operation, etc.). This type of error would occur in any
37implementation. In addition, a resource error may occur when a program exceeds the
38amount of available system resources (number of pending messages, system bu ers, etc.).
39The occurrence of this type of error depends on the amount of available resources in the
40system and the resource allocation mechanism used; this may di er from system to system.
41A high-quality implementation will provide generous limits on the important resources so
42as to alleviate the portability problem this represents.
43In C and Fortran, almost all MPI calls return a code that indicates successful completion
44of the operation. Whenever possible, MPI calls return an error code if an error occurred
45during the call. By default, an error detected during the execution of the MPI library
46causes the parallel computation to abort, except for le operations. However, MPI provides
47mechanisms for users to change this default and to handle recoverable errors. The user
48may specify that no error is fatal, and handle error codes returned by MPI calls by himself