- •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
16.2. FORTRAN SUPPORT |
485 |
real x
double precision d end type mytype
type(mytype) foo
integer blocklen(3), type(3) integer(MPI_ADDRESS_KIND) disp(3), base
call MPI_GET_ADDRESS(foo%i, disp(1), ierr) call MPI_GET_ADDRESS(foo%x, disp(2), ierr) call MPI_GET_ADDRESS(foo%d, disp(3), ierr)
base = disp(1)
disp(1) = disp(1) - base disp(2) = disp(2) - base disp(3) = disp(3) - base
blocklen(1) = 1 blocklen(2) = 1 blocklen(3) = 1
type(1) = MPI_INTEGER type(2) = MPI_REAL
type(3) = MPI_DOUBLE_PRECISION
call MPI_TYPE_CREATE_STRUCT(3, blocklen, disp, type, newtype, ierr) call MPI_TYPE_COMMIT(newtype, ierr)
!unpleasant to send foo%i instead of foo, but it works for scalar
!entities of type mytype
call MPI_SEND(foo%i, 1, newtype, ...)
A Problem with Register Optimization
MPI provides operations that may be hidden from the user code and run concurrently with it, accessing the same memory as user code. Examples include the data transfer for an MPI_IRECV. The optimizer of a compiler will assume that it can recognize periods when a copy of a variable can be kept in a register without reloading from or storing to memory. When the user code is working with a register copy of some variable while the hidden operation reads or writes the memory copy, problems occur. This section discusses register optimization pitfalls.
When a variable is local to a Fortran subroutine (i.e., not in a module or COMMON block), the compiler will assume that it cannot be modi ed by a called subroutine unless it is an actual argument of the call. In the most common linkage convention, the subroutine is expected to save and restore certain registers. Thus, the optimizer will assume that a register which held a valid copy of such a variable before the call will still hold a valid copy on return.
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
486 |
CHAPTER 16. LANGUAGE BINDINGS |
1Normally users are not a icted with this. But the user should pay attention to this
2section if in his/her program a bu er argument to an MPI_SEND, MPI_RECV etc., uses
3a name which hides the actual variables involved. MPI_BOTTOM with an MPI_Datatype
4containing absolute addresses is one example. Creating a datatype which uses one variable
5as an anchor and brings along others by using MPI_GET_ADDRESS to determine their
6o sets from the anchor is another. The anchor variable would be the only one mentioned in
7the call. Also attention must be paid if MPI operations are used that run in parallel with
8the user's application.
9Example 16.11 shows what Fortran compilers are allowed to do.
10
11
12
13
Example 16.11 Fortran 90 register optimization.
This source ... |
can be compiled as: |
call MPI_GET_ADDRESS(buf,bufaddr, |
call MPI_GET_ADDRESS(buf,...) |
|
14 |
|
|
ierror) |
|
|
15 |
call MPI_TYPE_CREATE_STRUCT(...) |
|
call MPI_TYPE_CREATE_STRUCT(1,1, |
||
16 |
|
|
bufaddr, |
|
|
17 |
|
|
MPI_REAL,type,ierror) |
|
|
18 |
call MPI_TYPE_COMMIT(...) |
|
call MPI_TYPE_COMMIT(type,ierror) |
||
19 |
register = buf |
|
val_old = buf |
||
20 |
val_old = register |
|
21 |
||
call MPI_RECV(MPI_BOTTOM,...) |
||
call MPI_RECV(MPI_BOTTOM,1,type,...) |
||
22 |
val_new = register |
|
val_new = buf |
The compiler does not invalidate the register because it cannot see that MPI_RECV changes the value of buf. The access of buf is hidden by the use of MPI_GET_ADDRESS
27
and MPI_BOTTOM.
28
Example 16.12 shows extreme, but allowed, possibilities.
29
30 Example 16.12 Fortran 90 register optimization { extreme.
31 |
Source |
compiled as |
or compiled as |
|
|
||||
32 |
call MPI_IRECV(buf,..req) |
call MPI_IRECV(buf,..req) |
call MPI_IRECV(buf,..req) |
|
33 |
||||
|
register = buf |
b1 = buf |
||
|
|
|||
34 |
call MPI_WAIT(req,..) |
call MPI_WAIT(req,..) |
call MPI_WAIT(req,..) |
|
|
||||
35 |
b1 = buf |
b1 := register |
|
|
|
|
|||
36 |
|
|
|
|
37 |
|
|
|
38MPI_WAIT on a concurrent thread modi es buf between the invocation of MPI_IRECV
39and the nish of MPI_WAIT. But the compiler cannot see any possibility that buf can be
40changed after MPI_IRECV has returned, and may schedule the load of buf earlier than
41typed in the source. It has no reason to avoid using a register to hold buf across the call to
42MPI_WAIT. It also may reorder the instructions as in the case on the right.
43To prevent instruction reordering or the allocation of a bu er in a register there are
44two possibilities in portable Fortran code:
45
46The compiler may be prevented from moving a reference to a bu er across a call to
47an MPI subroutine by surrounding the call by calls to an external subroutine with
48the bu er as an actual argument. Note that if the intent is declared in the external