text
based on the new Message-Passing Interface (MPI) standard.
After reading this book, students, scientists, and
engineers should be able to program any parallel system,
from networks of workstations to parallel supercomputers.
The only prerequisite is a knowledge of how to program
single-processor computers. Program examples in the text
are in C, and both C and
Foreword
Preface
Chapter 1
Introduction
1.1 The Need for More
Computational Power
1.2 The Need for Parallel Computing
1.3 The Bad News
1.4 MPI
1.5 The Rest of the Book
1.6 Typographic Conventions
Chapter 2 An Overview of
Parallel Computing
2.1 Hardware
2.1.1 Flynn's Taxonomy
2.1.2 The Classical von Neumann Machine
2.1.3 Pipeline and Vector Architectures
2.1.4 SIMD Systems
2.1.5 General MIMD Systems
2.1.6 Shared-Memory MIMD
2.1.7 Distributed-Memory MIMD
2.1.8
Communication and Routing
2.2 Software Issues
2.2.1 Shared-Memory
Programming
2.2.2 Message Passing
2.2.3 Data-Parallel Languages
2.2.4 RPC and Active Messages
2.2.5 Data Mapping
2.3 Summary
2.4 References
2.5 Exercises
3.1 The Program
3.2 Execution
3.3 MPI
3.3.1 General MPI
Programs
3.3.2 Finding Out about the Rest of the World
3.3.3 Message: Data + Envelope
3.3.4 Sending Messages
3.4 Summary
3.5 References
3.6 Exercises
3.7 Programming Assignment
4.1 The Trapezoidal
Rule
4.2 Parallelizing the Trapezoidal Rule
4.3 I/O on Parallel Systems
4.4 Summary
4.5 References
4.6 Exercises
4.7 Programming Assignments
Chapter 5 Collective
Communication
5.1 Tree-Structured
Communication
5.2 Broadcast
5.3 Tags, Safety, Buffering, and Synchronization
5.4 Reduce
5.5 Dot Product
5.6 Allreduce
5.7 Gather and Scatter
5.8 Summary
5.10 References
5.11 Exercises
5.12 Programming Assignments
Chapter 6 Grouping Data for
Communication
6.1 The Count Parameter
6.2 Derived Types and MPI_Type_struct
6.3 Other Derived Datatype Constructors
6.4 Type Matching
6.5 Pack/Unpack
6.6 Deciding Which Method to Use
6.7 Summary
6.8 References
6.9 Exercises
6.10 Programming Assignments
Chapter 7 Communicators and
Topologies
7.1 Matrix
Multiplication
7.2 Fox's Algorithm
7.3 Communicators
7.4 Working with Groups, Contexts, and Communicators
7.5 MPI_Comm_split
7.6 Topologies
7.7 MPI_Cart_sub
7.8 Implementation of Fox's Algorithm
7.9 Summary
7.10 References
7.11 Exercises
7.12 Programming Assignments
Chapter 8 Dealing with
I/O
8.1 Dealing with stdin,
stdout, and stderr
8.1.1 Attribute caching
8.1.2 Callback Functions
8.1.3 Identifying the I/O process rank
8.1.4 Caching an I/O Process Rank
8.1.5 Retrieving the I/O Process Rank
8.1.6 Reading from stdin
8.1.7 Writing to stdout
8.1.8 Writing to stderr and Error Checking
8.2 Limited Access to stdin
8.3 File I/O
8.4 Array I/O
8.4.1 Data
Distributions
8.4.2 Model Problem
8.4.3 Distribution of the Input
8.4.4 Derived Datatypes
8.4.5 The Extent of a Derived Datatype
8.4.6 The Input Code
8.4.7 Printing the Array
8.4.8 An Example
8.5 Summary
8.6 References
8.7 Exercises
8.8 Programming Assignments
9.1 Quick Review of Serial
Debugging
9.1.1 Examine the Source
Code
9.1.2 Add Debugging Output
9.1.3 Use a Debugger
9.2 More on Serial debugging
9.3 Parallel Debugging
9.4 Nondeterminism
9.5 An Example
9.5.1 The Program?
9.5.2 Debugging The Program
9.5.3 A Brief Discussion of Parallel Debuggers
9.5.4 The Old Standby: printf/fflush
9.5.5 The Classical Bugs in Parallel Programs
9.5.6 First Fix
9.5.7 many parallel Programming Bugs are Really Serial
Programming Bugs
9.5.8 Different Systems, Different Errors
9.5.9 Moving to Multiple Processes
9.5.10 Confusion about I/O
9.5.11 Finishing Up
9.6 Error Handling in MPI
9.7 Summary
9.8 References
9.9 Exercises
9.10 Programming Assignments
10.1 Data-Parallel
Programs
10.2 Jacobi's Method
10.3 Parallel Jacobi's Method
10.4 Coding Parallel Programs
10.5 An Example: Sorting
10.5.1 Main Program
10.5.2 The "Input" Functions
10.5.3 All-to-all Scatter/Gather
10.5.4 Redistributing the Keys
10.5.5 Pause to Clean Up
10.5.6 Find_alltoall_send_params
10.5.7 Finishing Up
10.8 Summary
10.7 References
10.8 Exercises
10.9 Programming Assignments
11.1 Serial Program
Performance
11.2 An Example: The Serial Trapezoidal Rule
11.3 What about the I/O?
11.4 Parallel Program Performance Analysis
11.5 The Cost of Communication
11.6 An Example: The Parallel Trapezoidal Rule
11.7 Taking Timings
11.8 Summary
11.9 References
11.10 Exercises
11.11 Programming Assignments
Chapter 12 More on
Performance
12.1 Amdahl's Law
12.2 Work and Overhead
12.3 Sources of Overhead
12.4 Scalability
12.5 Potential Problems in Estimating Performance
12.5.1 Networks of
Workstations and Resource Contention
12.5.2 Load Balancing and Idle Time
12.5.3 Overlapping Communication and Computation
12.5.4 Collective Communication 12.6 Performance Evaluation
Tools
12.6.1 MPI's Profiling
Interface
12.6.2 Upshot
12.7 Summary
12.8 References
12.9 Exercises
12.10 Programming Assignments
13.1 An Example: Coding
Allgather
13.1.1 Function
Parameters
13.1.2 Ring Pass Allgather
13.2 Hypercubes
13.2.1 Additional Issues in
the Hypercube Exchange
13.2.2 Details of the Hypercube Algorithm
13.3 Send-receive
13.4 Null Processes
13.5 Nonblocking Communication
13.5.1 Ring Allgather with
Nonblocking Communication
13.5.2 Hypercube Allgather with Nonblocking
Communication
13.6 Persistent Communication Requests
13.7 Communication Modes
13.7.1 Synchronous Mode
13.7.2 Ready Mode
13.7.3 Buffered Mode
13.8 The Last Word on Point-to-Point Communication
13.9 Summary
13.10 References
13.11 Exercises
13.12 Programming Assignments
14.1 Designing a Parallel
Algorithm
14.2 Sorting
14.3 Serial Bitonic Sort
14.4 Parallel Bitonic Sort
14.5 Tree Searches and Combinatorial Optimization
14.6 Serial Tree Search
14.7 Parallel Tree Search
14.7.1 Par_dfs
14.7.2 Service_requests
14.7.3 Work_remains
14.7.4 Distributed Termination Detection
14.8 Summary
14.9 References
14.10 Exercises
14.11 Programming Assignments
15.1 Using Libraries: Pro and
Con
15.2 Using More than One Language
15.3 ScaLAPACK
15.4 An Example of a ScaLAPACK Program
15.5 PETSc
15.6 A PETSc Example
15.7 Summary
15.8 References
15.9 Exercises
15.10 Programming Assignments
Chapter 16 Wrapping
Up
16.1 Where to Go from
Here
16.2 The Future of MPI
Appendix A Summary of MPI
Commands
A.1 Point-to-Point
Communication Functions
A.1.1 Blocking Sends and
Receives
A.1.2 Communication Modes
A.1.3 Buffer Allocation
A.1.4 Nonblocking Communication
A.1.5 Probe and Cancel
A.1.6 Persistent Communication Requests
A.1.7 Send-receive
A.2 Derived Datatypes and MPI_Pack/Unpack
A.2.1 Derived Datatypes
A.2.2 MPI_Pack and MPI_Unpack
A.3 Collective Communication Functions
A.3.1 Barrier and
Broadcast
A.3.2 Gather Scatter
A.3.3 Reduction Operations
A.4 Groups, Contexts, and Communicators
A.4.1 Group Management
A.4.2 Communicator Management
A.4.3 Inter-communicators
A.4.4 Attribute Caching
A.5 Process Topologies
A.5.1 General Topology
Functions
A.5.2 Cartesian Topology Management
A.5.3 Graph Topology Management
A.6 Environmental Management
A.6.1 Implementation
Information
A.6.2 Error Handling
A.6.3 Timers
A.6.4 Startup
A.7 Profiling
A.8 Constants
A.9 Type Definitions
B.1 Implementations of
MPI
B.2 The MPI FAQ
B.3 MPI Web Pages
B.4 MPI Newsgroup
B.5 MPI-2 and MPI-IO
B.6 Parallel Programming with MPI
Bibliography
Index