- •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
306 |
CHAPTER 10. PROCESS CREATION AND MANAGEMENT |
clean interface between an application and system software.
3MPI must guarantee communication determinism in the presense of dynamic processes,
4i.e., dynamic process management must not introduce unavoidable race conditions.
5 |
MPI must not contain features that compromise performance. |
6 |
|
7 |
The process management model addresses these issues in two ways. First, MPI remains |
|
|
8 |
primarily a communication library. It does not manage the parallel environment in which |
|
|
9 |
a parallel program executes, though it provides a minimal interface between an application |
|
|
10 |
and external resource and process managers. |
|
|
11 |
Second, MPI maintains a consistent concept of a communicator, regardless of how its |
|
|
12 |
members came into existence. A communicator is never changed once created, and it is |
|
|
13 |
always created using deterministic collective operations. |
|
|
14 |
|
15 |
|
16 |
10.2 The Dynamic Process Model |
17 |
|
18The dynamic process model allows for the creation and cooperative termination of processes
19after an MPI application has started. It provides a mechanism to establish communication
20between the newly created processes and the existing MPI application. It also provides a
21mechanism to establish communication between two existing MPI applications, even when
22one did not \start" the other.
10.2.1 Starting Processes
MPI applications may start new processes through an interface to an external process manager.
MPI_COMM_SPAWN starts MPI processes and establishes communication with them, returning an intercommunicator. MPI_COMM_SPAWN_MULTIPLE starts several di erent binaries (or the same binary with di erent arguments), placing them in the same MPI_COMM_WORLD and returning an intercommunicator.
MPI uses the existing group abstraction to represent processes. A process is identi ed
32
by a (group, rank) pair.
33
34
10.2.2 The Runtime Environment
35
36The MPI_COMM_SPAWN and MPI_COMM_SPAWN_MULTIPLE routines provide an inter-
37face between MPI and the runtime environment of an MPI application. The di culty is that
38there is an enormous range of runtime environments and application requirements, and MPI
39must not be tailored to any particular one. Examples of such environments are:
40
41MPP managed by a batch queueing system. Batch queueing systems generally
42allocate resources before an application begins, enforce limits on resource use (CPU
43time, memory use, etc.), and do not allow a change in resource allocation after a
44job begins. Moreover, many MPPs have special limitations or extensions, such as a
45limit on the number of processes that may run on one processor, or the ability to
46gang-schedule processes of a parallel application.
47
48
10.2. THE DYNAMIC PROCESS MODEL |
307 |
Network of workstations with PVM. PVM (Parallel Virtual Machine) allows a user to create a \virtual machine" out of a network of workstations. An application may extend the virtual machine or manage processes (create, kill, redirect output, etc.) through the PVM library. Requests to manage the machine or processes may be intercepted and handled by an external resource manager.
Network of workstations managed by a load balancing system. A load balancing system may choose the location of spawned processes based on dynamic quantities, such as load average. It may transparently migrate processes from one machine to another when a resource becomes unavailable.
Large SMP with Unix. Applications are run directly by the user. They are scheduled at a low level by the operating system. Processes may have special scheduling characteristics (gang-scheduling, processor a nity, deadline scheduling, processor locking, etc.) and be subject to OS resource limits (number of processes, amount of memory, etc.).
MPI assumes, implicitly, the existence of an environment in which an application runs. It does not provide \operating system" services, such as a general ability to query what processes are running, to kill arbitrary processes, to nd out properties of the runtime environment (how many processors, how much memory, etc.).
Complex interaction of an MPI application with its runtime environment should be done through an environment-speci c API. An example of such an API would be the PVM task and machine management routines | pvm_addhosts, pvm_config, pvm_tasks, etc., possibly modi ed to return an MPI (group,rank) when possible. A Condor or PBS API would be another possibility.
At some low level, obviously, MPI must be able to interact with the runtime system, but the interaction is not visible at the application level and the details of the interaction are not speci ed by the MPI standard.
In many cases, it is impossible to keep environment-speci c information out of the MPI interface without seriously compromising MPI functionality. To permit applications to take advantage of environment-speci c functionality, many MPI routines take an info argument that allows an application to specify environment-speci c information. There is a tradeo between functionality and portability: applications that make use of info are not portable.
MPI does not require the existence of an underlying \virtual machine" model, in which there is a consistent global view of an MPI application and an implicit \operating system" managing resources and processes. For instance, processes spawned by one task may not be visible to another; additional hosts added to the runtime environment by one process may not be visible in another process; tasks spawned by di erent processes may not be automatically distributed over available resources.
Interaction between MPI and the runtime environment is limited to the following areas:
A process may start new processes with MPI_COMM_SPAWN and
MPI_COMM_SPAWN_MULTIPLE.
When a process spawns a child process, it may optionally use an info argument to tell the runtime environment where or how to start the process. This extra information may be opaque to MPI.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48