MPI: The Complete Reference (Vol. 1) - 2nd Edition

by Marc Snir, Steve Otto, Steven Huss-Lederman, David Walker and Jack Dongarra

• ISBN13: 9780262692151
• ISBN10: 0262692155
• Edition: 2nd
• Format: Paperback
• Copyright: 1998-09-19
• Publisher: Mit Pr

Summary

Since its release in summer 1994, the Message Passing Interface (MPI) specification has become a standard for message-passing libraries for parallel computations. There exist more than a dozen implementations on a variety of computing platforms, from the IBM SP-2 supercomputer to PCs running Windows NT. The initial MPI Standard, known as MPI-1, has been modified over the last two years. This volume, the definitive reference manual for the latest version of MPI-1, contains a complete specification of the MPI Standard. It is annotated with comments that clarify complicated issues, including why certain design choices were made, how users are intended to use the interface, and how they should construct their version of MPI. The volume also provides many detailed, illustrative programming examples.

Table of Contents

Series Foreword	xiii	(2)
Preface	xv
1 Introduction	1	(28)
1.1 Message Passing and Standardization	1	(3)
1.2 Who Should Use This Standard?	4	(1)
1.3 What Platforms Are Targets for Implementation?	4	(1)
1.4 What Is Included in MPI?	5	(1)
1.5 Version of MPI	6	(1)
1.6 MPI Conventions and Design Choices	6	(3)
1.6.1 Document Notation	6	(1)
1.6.2 Naming Conventions	7	(1)
1.6.3 Procedure Specification	8	(1)
1.7 Semantic Terms	9	(2)
1.8 Function Argument Data Type	11	(3)
1.8.1 Opaque Objects	11	(2)
1.8.2 Array Arguments	13	(1)
1.8.3 State	13	(1)
1.8.4 Named Constants	13	(1)
1.8.5 Choice	14	(1)
1.8.6 Addresses	14	(1)
1.8.7 File Offsets	14	(1)
1.9 Language Binding	14	(9)
1.9.1 Deprecated Names and Functions	15	(1)
1.9.2 Fortran Binding Issues	15	(3)
1.9.3 C Binding Issues	18	(1)
1.9.4 C++ Binding Issues	19	(4)
1.10 Processes	23	(1)
1.11 Error Handling	24	(2)
1.12 Implementation Issues	26	(3)
1.12.1 Independence of Basic Runtime Routines	26	(1)
1.12.2 Interaction with Signals	27	(2)
2 Point-to-Point Communication	29	(94)
2.1 Introduction and Overview	29	(4)
2.2 Blocking Send and Receive Operations	33	(11)
2.2.1 Blocking Send	33	(1)
2.2.2 Send Buffer and Message Data	33	(3)
2.2.3 Message Envelope	36	(1)
2.2.4 Comments on Send	37	(1)
2.2.5 Blocking Receive	38	(1)
2.2.6 Receive Buffer	39	(1)
2.2.7 Message Selection	39	(1)
2.2.8 Return Status	39	(3)
2.2.9 Comments on Receive	42	(2)
2.3 Datatype Matching and Data Conversion	44	(5)
2.3.1 Type Matching Rules	44	(3)
2.3.2 Data Conversion	47	(2)
2.3.3 Comments on Data Conversion	49	(1)
2.4 Semantics of Blocking Point-to-Point	49	(7)
2.4.1 Buffering and Safety	49	(3)
2.4.2 Order	52	(3)
2.4.3 Progress	55	(1)
2.4.4 Fairness	55	(1)
2.5 Example--Jacobi Iteration	56	(5)
2.6 Send-Receive	61	(4)
2.7 Null Processes	65	(3)
2.8 Nonblocking Communication	68	(20)
2.8.1 Request Objects	69	(1)
2.8.2 Posting Operations	69	(2)
2.8.3 Completion Operations	71	(2)
2.8.4 Examples	73	(5)
2.8.5 Freeing Requests	78	(1)
2.8.6 Nondestructive Test of status	79	(1)
2.8.7 Semantics of Nonblocking Communications	80	(5)
2.8.8 Comments on Semantics of Nonblocking Communications	85	(3)
2.9 Multiple Completions	88	(9)
2.10 Probe and Cancel	97	(6)
2.11 Persistent Communication Requests	103	(5)
2.12 Communication-Complete Calls with Null Request Handles	108	(3)
2.13 Communication Modes	111	(12)
2.13.1 Blocking Calls	113	(1)
2.13.2 Nonblocking Calls	114	(2)
2.13.3 Persistent Requests	116	(2)
2.13.4 Buffer Allocation and Usage	118	(2)
2.13.5 Model Implementation of Buffered Mode	120	(1)
2.13.6 Comments on Communication Modes	121	(2)
3 User-Defined Datatypes and Packing	123	(68)
3.1 Introduction	123	(1)
3.2 Introduction to User-Defined Datatypes	123	(4)
3.3 Datatype Accessors	127	(1)
3.4 Datatype Constructors	128	(19)
3.4.1 Dup	128	(1)
3.4.2 Contiguous	129	(1)
3.4.3 Vector	130	(2)
3.4.4 Hvector	132	(5)
3.4.5 Indexed	137	(2)
3.4.6 Block Indexed	139	(1)
3.4.7 Hindexed	140	(2)
3.4.8 Struct	142	(5)
3.5 Use of Derived Datatypes	147	(5)
3.5.1 Commit	147	(1)
3.5.2 Deallocation	148	(1)
3.5.3 Relation to count	149	(1)
3.5.4 Type Matching	149	(1)
3.5.5 Message Length	150	(2)
3.6 Address Function	152	(2)
3.7 Datatype Resizing	154	(6)
3.7.1 True Extent of Datatypes	159	(1)
3.8 Absolute Addresses	160	(1)
3.9 Array Datatype Constructors	161	(9)
3.10 Portability of Datatypes	170	(1)
3.11 Deprecated Functions	171	(4)
3.12 Pack and Unpack	175	(16)
3.12.1 Canonical MPI_PACK and MPI_UNPACK	184	(4)
3.12.2 Derived Datatypes versus Pack/Unpack	188	(3)
4 Collective Communications	191	(64)
4.1 Introduction and Overview	191	(2)
4.2 Operational Details	193	(1)
4.3 Communicator Argument	194	(1)
4.4 Barrier Synchronization	195	(1)
4.5 Global Communication Functions	195	(2)
4.6 Broadcast	197	(2)
4.6.1 An Example Using MPI_BCAST	198	(1)
4.7 Gather	199	(11)
4.7.1 Examples Using MPI_GATHER	200	(2)
4.7.2 Gather, Vector Variant	202	(1)
4.7.3 Examples Using MPI_GATHERV	203	(7)
4.8 Scatter	210	(5)
4.8.1 An Example Using MPI_SCATTER	211	(1)
4.8.2 Scatter: Vector Variant	212	(1)
4.8.3 Examples Using MPI_SCATTERV	213	(2)
4.9 Gather to All	215	(3)
4.9.1 An Example Using MPI_ALLGATHER	216	(1)
4.9.2 Gather to All: Vector Variant	217	(1)
4.10 All to All Scatter/Gather	218	(7)
4.10.1 All to All: Vector Variant	219	(2)
4.10.2 All to All: Generalized Function	221	(1)
4.10.3 An Example Using MPI_ALLTOALLW	222	(3)
4.11 Global Reduction Operations	225	(17)
4.11.1 Reduce	225	(4)
4.11.2 Predefined Reduce Operations	229	(3)
4.11.3 MINLOC and MAXLOC	232	(6)
4.11.4 All Reduce	238	(2)
4.11.5 Reduce-Scatter	240	(2)
4.12 Scan Operations	242	(2)
4.12.1 Inclusive Scan	242	(1)
4.12.2 Exclusive Scan	243	(1)
4.13 User-Defined Operations for Reduce and Scan	244	(6)
4.14 The Semantics of Collective Communications	250	(5)
5 Communicators	255	(64)
5.1 Introduction	255	(2)
5.1.1 Division of Processes	255	(1)
5.1.2 Avoiding Message Conflicts between Modules	256	(1)
5.1.3 Extensibility by Users	256	(1)
5.1.4 Safety	256	(1)
5.2 Overview	257	(5)
5.2.1 Groups	257	(1)
5.2.2 Communicator	258	(1)
5.2.3 Communication Domains	259	(2)
5.2.4 Compatibility with Previous Practice	261	(1)
5.3 Group Management	262	(10)
5.3.1 Group Accessors	262	(2)
5.3.2 Group Constructors	264	(7)
5.3.3 Group Destructors	271	(1)
5.4 Communicator Management	272	(8)
5.4.1 Communicator Accessors	272	(2)
5.4.2 Communicator Constructors	274	(5)
5.4.3 Communicator Destructor	279	(1)
5.5 Safe Parallel Libraries	280	(7)
5.6 Caching	287	(17)
5.6.1 Introduction	287	(3)
5.6.2 Caching Functions	290	(11)
5.6.3 Deprecated Functions	301	(3)
5.7 Intercommunication	304	(15)
5.7.1 Introduction	304	(4)
5.7.2 Intercommunicator Accessors	308	(2)
5.7.3 Intercommunicator Constructors and Destructors	310	(4)
5.7.4 Examples	314	(5)
6 Process Topologies	319	(34)
6.1 Introduction	319	(1)
6.2 Virtual Topologies	320	(1)
6.3 Overlapping Topologies	321	(1)
6.4 Embedding in MPI	322	(1)
6.5 Cartesian Topology Functions	323	(10)
6.5.1 Cartesian Constructor Function	323	(1)
6.5.2 Cartesian Convenience Function: MPI_DIMS_CREATE	324	(1)
6.5.3 Cartesian Inquiry Functions	325	(2)
6.5.4 Cartesian Translator Functions	327	(1)
6.5.5 Cartesian Shift Function	328	(3)
6.5.6 Cartesian Partition Function	331	(1)
6.5.7 Cartesian Low-Level Functions	332	(1)
6.6 Graph Topology Functions	333	(7)
6.6.1 Graph Constructor Function	334	(2)
6.6.2 Graph Inquiry Functions	336	(1)
6.6.3 Graph Information Functions	337	(2)
6.6.4 Low-Level Graph Functions	339	(1)
6.7 Topology Inquiry Functions	340	(1)
6.8 An Application Example	340	(13)
7 Environmental Management	353	(34)
7.1 MPI Process Startup	353	(4)
7.2 Intialization and Exit	357	(7)
7.3 Implementation Information	364	(4)
7.3.1 MPI Version Number	364	(1)
7.3.2 Environmental Inquiries	365	(3)
7.4 Timers and Synchronization	368	(1)
7.5 Error Handling	369	(15)
7.5.1 Error Handlers	371	(8)
7.5.2 Deprecated Functions	379	(1)
7.5.3 Error Codes	380	(4)
7.6 Interaction with Executing Environment	384	(3)
7.6.1 Independence of Basic Runtime Routines	384	(1)
7.6.2 Interaction with Signals in POSIX	385	(2)
8 The MPI Profiling Interface	387	(8)
8.1 Requirements	387	(1)
8.2 Discussion	388	(1)
8.3 Logic of the Design	388	(2)
8.3.1 Miscellaneous Control of Profiling	389	(1)
8.4 Examples	390	(4)
8.4.1 Profiler Implementation	390	(1)
8.4.2 MPI Library Implementation	391	(2)
8.4.3 Complications	393	(1)
8.5 Multiple Levels of Interception	394	(1)
9 Conclusions	395	(14)
9.1 Design Issues	395	(2)
9.1.1 Why Is MPI So Big?	395	(1)
9.1.2 Should We Be Concerned about the Size of MPI?	395	(1)
9.1.3 Why Does MPI Not Guarantee Buffering?	396	(1)
9.2 Portable Programming with MPI	397	(9)
9.2.1 Dependency on Buffering	398	(6)
9.2.2 Collective Communication and Synchronization	404	(1)
9.2.3 Ambiguous Communications and Portability	404	(2)
9.3 Heterogeneous Computing with MPI	406	(1)
9.4 MPI Implementations	407	(2)
References	409	(2)
Constants Index	411	(4)
Function Index	415	(6)
Index	421