- •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
5.10. REDUCE-SCATTER |
177 |
MPI_REDUCE_LOCAL( inbuf, inoutbuf, count, datatype, op)
IN |
inbuf |
input bu er (choice) |
INOUT |
inoutbuf |
combined input and output bu er (choice) |
IN |
count |
number of elements in inbuf and inoutbuf bu ers (non- |
|
|
negative integer) |
IN |
datatype |
data type of elements of inbuf and inoutbuf bu ers |
|
|
(handle) |
IN |
op |
operation (handle) |
int MPI_Reduce_local(void* inbuf, void* inoutbuf, int count, MPI_Datatype datatype, MPI_Op op)
MPI_REDUCE_LOCAL(INBUF, INOUBUF, COUNT, DATATYPE, OP, IERROR) <type> INBUF(*), INOUTBUF(*)
INTEGER COUNT, DATATYPE, OP, IERROR
fvoid MPI::Op::Reduce_local(const void* inbuf, void* inoutbuf, int count, const MPI::Datatype& datatype) const (binding deprecated, see Section 15.2) g
The function applies the operation given by op element-wise to the elements of inbuf and inoutbuf with the result stored element-wise in inoutbuf, as explained for user-de ned operations in Section 5.9.5. Both inbuf and inoutbuf (input as well as result) have the same number of elements given by count and the same datatype given by datatype. The MPI_IN_PLACE option is not allowed.
Reduction operations can be queried for their commutativity.
MPI_OP_COMMUTATIVE( op, commute)
IN |
op |
operation (handle) |
OUT |
commute |
true if op is commutative, false otherwise (logical) |
int MPI_Op_commutative(MPI_Op op, int *commute)
MPI_OP_COMMUTATIVE(OP, COMMUTE, IERROR)
LOGICAL COMMUTE
INTEGER OP, IERROR
fbool MPI::Op::Is_commutative() const (binding deprecated, see Section 15.2) g
5.10 Reduce-Scatter
MPI includes variants of the reduce operations where the result is scattered to all processes in a group on return. One variant scatters equal-sized blocks to all processes, while another variant scatters blocks that may vary in size for each process.
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
1
2
3
4
5
178 |
CHAPTER 5. COLLECTIVE COMMUNICATION |
5.10.1 MPI_REDUCE_SCATTER_BLOCK
MPI_REDUCE_SCATTER_BLOCK( sendbuf, recvbuf, recvcount, datatype, op, comm)
6 |
IN |
sendbuf |
starting address of send bu er (choice) |
7 |
OUT |
recvbuf |
starting address of receive bu er (choice) |
8 |
IN |
recvcount |
element count per block (non-negative integer) |
|
|||
9 |
|
datatype |
data type of elements of send and receive bu ers (han- |
10 |
IN |
||
|
|
dle) |
|
11 |
|
|
|
|
|
|
|
12 |
IN |
op |
operation (handle) |
13 |
IN |
comm |
communicator (handle) |
|
|||
14 |
|
|
|
15 |
int MPI_Reduce_scatter_block(void* sendbuf, void* recvbuf, int recvcount, |
||
|
|||
16 |
|
MPI_Datatype datatype, MPI_Op op, MPI_Comm comm) |
|
|
|
||
17 |
|
|
|
18 |
MPI_REDUCE_SCATTER_BLOCK(SENDBUF, RECVBUF, RECVCOUNT, DATATYPE, OP, COMM, |
||
19 |
|
IERROR) |
|
20 |
<type> SENDBUF(*), RECVBUF(*) |
|
|
21 |
INTEGER RECVCOUNT, DATATYPE, OP, COMM, IERROR |
||
22 |
fvoid MPI::Comm::Reduce_scatter_block(const void* sendbuf, void* recvbuf, |
||
23 |
|||
int recvcount, const MPI::Datatype& datatype,
24 |
(binding deprecated, see Section 15.2) g |
const MPI::Op& op) const = 0 |
|
25 |
|
26If comm is an intracommunicator, MPI_REDUCE_SCATTER_BLOCK rst performs a
27global, element-wise reduction on vectors of count = n*recvcount elements in the send bu ers
28de ned by sendbuf, count and datatype, using the operation op, where n is the number of
29processes in the group of comm. The routine is called by all group members using the
30same arguments for recvcount, datatype, op and comm. The resulting vector is treated as
31n consecutive blocks of recvcount elements that are scattered to the processes of the group.
32The i-th block is sent to process i and stored in the receive bu er de ned by recvbuf,
33recvcount, and datatype.
34
35
36
37
38
Advice to implementors. The MPI_REDUCE_SCATTER_BLOCK routine is functionally equivalent to: an MPI_REDUCE collective operation with count equal to recvcount*n, followed by an MPI_SCATTER with sendcount equal to recvcount. However, a direct implementation may run faster. (End of advice to implementors.)
39The \in place" option for intracommunictors is speci ed by passing MPI_IN_PLACE in
40the sendbuf argument on all processes. In this case, the input data is taken from the receive
41bu er.
42If comm is an intercommunicator, then the result of the reduction of the data provided
43by processes in one group (group A) is scattered among processes in the other group (group
44B) and vice versa. Within each group, all processes provide the same value for the recvcount
45argument, and provide input vectors of count = n*recvcount elements stored in the send
46bu ers, where n is the size of the group. The number of elements count must be the same
47for the two groups. The resulting vector from the other group is scattered in blocks of
48recvcount elements among the processes in the group.
5.10. REDUCE-SCATTER |
179 |
Rationale. The last restriction is needed so that the length of the send bu er of one group can be determined by the local recvcount argument of the other group. Otherwise, a communication is needed to gure out how many elements are reduced. (End of rationale.)
5.10.2 MPI_REDUCE_SCATTER
MPI_REDUCE_SCATTER extends the functionality of MPI_REDUCE_SCATTER_BLOCK such that the scattered blocks can vary in size. Block sizes are determined by the recvcounts array, such that the i-th block contains recvcounts[i] elements.
MPI_REDUCE_SCATTER( sendbuf, recvbuf, recvcounts, datatype, op, comm)
IN |
sendbuf |
starting address of send bu er (choice) |
OUT |
recvbuf |
starting address of receive bu er (choice) |
IN |
recvcounts |
non-negative integer array (of length group size) spec- |
|
|
ifying the number of elements of the result distributed |
|
|
to each process. |
IN |
datatype |
data type of elements of send and receive bu ers (han- |
|
|
dle) |
IN |
op |
operation (handle) |
IN |
comm |
communicator (handle) |
int MPI_Reduce_scatter(void* sendbuf, void* recvbuf, int *recvcounts, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
MPI_REDUCE_SCATTER(SENDBUF, RECVBUF, RECVCOUNTS, DATATYPE, OP, COMM, IERROR)
<type> SENDBUF(*), RECVBUF(*)
INTEGER RECVCOUNTS(*), DATATYPE, OP, COMM, IERROR
fvoid MPI::Comm::Reduce_scatter(const void* sendbuf, void* recvbuf, int recvcounts[], const MPI::Datatype& datatype,
const MPI::Op& op) const = 0 (binding deprecated, see Section 15.2) g
If comm is an intracommunicator, MPI_REDUCE_SCATTER rst performs a global,
element-wise reduction on vectors of count = Pn 1 recvcounts[i] elements in the send bu ers
i=0
de ned by sendbuf, count and datatype, using the operation op, where n is the number of processes in the group of comm. The routine is called by all group members using the same arguments for recvcounts, datatype, op and comm. The resulting vector is treated as n consecutive blocks where the number of elements of the i-th block is recvcounts[i]. The blocks are scattered to the processes of the group. The i-th block is sent to process i and stored in the receive bu er de ned by recvbuf, recvcounts[i] and datatype.
Advice to implementors. The MPI_REDUCE_SCATTER routine is functionally equivalent to: an MPI_REDUCE collective operation with count equal to the sum of recvcounts[i] followed by MPI_SCATTERV with sendcounts equal to recvcounts. However, a direct implementation may run faster. (End of advice to implementors.)
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