Using OpenMP: Portable Shared Memory Parallel Programming (Scientific and Engineering Computation) [Paperback]

I have most of the parallel computing books out there so I am sort of a collector of sorts. Most focus on either the basics of parallel programming, MPI, OpenMP, both, or some other less popular (yet) paradigm e.g. PFortran, TBB, etc. With every parallel-computing wanna be buying a multicore machine dual, quad, dual-quad, etc., the parallel computing software "industry" is in flux. No longer will MPI on a cluster be enough. It still remains to be seen whether the slower memory bus on quad core machines will allow for speedups without major code overhaul or a new paradigm. Anyway, this book is a welcome addition to my collection. For one, it is current e.g. 2008 and also it is focussed on OpenMP (but does treat dual MPI/OpenMP programming). It is well written (I am about 100 pages in since I just got my copy last week) and has one tantalizing chapter entitled "How to get good performance by using OpenMP" - which is really timely since my new 72 core machine (9 dual Intel quad cores) seems to give slower performance for a major commercial CFD code than the equivalent number of dual-core nodes). I hope it helps me. Based on the rapid growth of multicore machines and the lack of a simple programming solution, I recommend this book to all those wanting to try and get their codes running fast on multicore machines. The only downsides in this book so far is the lack of downloadable code (you have to type it in yourself) and it is hard to test the code fragments because they are just that - fragments. A nice feature of the book is the 50/50 emphasis on Fortran/C codes - which are the still the mainstay in large-scale scientific computing.

The OpenMP specification can be downloaded from the web, but it is not a really a good starting point for learning how to write real programs using the OpenMP constructs. However, this book does have a lot of material that you really don't need just to write programs. This extra information is in the form of context and information on parallel computing in general, since this book is really intended to double as a textbook and a practical guide for professionals. The following briefly describes the contents.

Chapter one contains some background information on OpenMP and its applications. You can skip it if you are not interested in this or already know the material.
Chapter two is a brief overview of the features of OpenMP at a high level. It discusses how OpenMP deals with problems that come from the complex memory hierarchy that exists on modern computers.
Chapter three is a good starting point if you know you need OpenMP, know why you need it, and just need to get something going. It discusses a complete OpenMP program in both C and Fortran that uses a couple of OpenMP's most widely used features, plus it explains the basics of the OpenMP syntax. The problem discussed is specifically how to perform a matrix times a vector operation in parallel.
Chapter four is a more complete overview of the OpenMP programming paradigm and it contains many example programs. First the most widely used features are introduced with a focus on those features that enable work to be shared among multiple threads. The scope narrows until the author is down to some of OpenMP's less widely known features. The programs start simple and get more complex as the chapter progresses, always staying within the field of scientific computing.
Chapters five and six go together, and discuss how to optimize performance with OpenMP. There are a number of practical programming tips and an extended example that gives insight into the process of investigating performance problems.
Chapter seven talks about program correctness and troubleshooting. This can be hard to do in shared-memory parallel programs.
Chapter 8 is on translation by the compiler of an OpenMP program into an application that can be executed in parallel. It talks about behind the scenes occurrences with OpenMP computing including the operation of OpenMP-aware compilers, performance tools, and debuggers. It also discusses strategies for obtaining high performance.
Chapter 9 is a special topics chapter and discusses trends that could influence extensions to the OpenMP specification in the future. Obviously it is not necessary for the practicing professional to know this, but it is interesting.

My perspective is that of someone that knows I must use OpenMP, and I need good concrete examples and an accompanying tutorial to get me going. Chapters three through eight were ideal for my purpose. Other books I examined ran the gamut from talking about why OpenMP is important and lacking practical details, to overpriced textbooks, to books that were about OpenMP plus some other parallel programming paradigms and weren't specific or modern enough. This one is clear, concise, modern, and the price can't be beat. The only drawback for me was the dual emphasis on C and Fortran, but then I'm sure the Fortran information is still useful to a good number of programmers.

Only the most inexpensive processors, or processors built for low power consumption, now have single cores. The present and future of CPUs is multi-core, quad cores per CPU, 6 cores soon to come, and probably more after that. The authors have a deep understanding of parallel processing, modern computer architecture, and OpenMP. This understanding is communicated clearly in this excellent book. The only reason to use OpenMP is to make your programs run faster, this motivation permeates the entire book. Extensive discussions regarding performance are included, including extensive discussions of coding to maximize hits on the CPU cache, considerations of overhead in parallel program, how memory placement and thread binding behavior of multiple multi-core CPUs can affect performance, and many other considerations that likely never occured to you. Almost all of the discussions are presented with specific examples and instruction regarding how to code OpenMP directives. The emphasis is on C, with enough examples in Fortran to be able to use that also -- there is no discussion of C++. Since C and Fortran are by far the most important languages used for scientific computation, the language choices are appropriate at least for that community.