Writing Scientific Software: A Guide to Good Style [Paperback]

I think this book is a good introduction to writing good/better scientific programs. It seems so very much is left out of courses, and people are left to their own devices in terms of learning how to write good, robust, and fast programs. This book is a mix of numerical analysis, programming techniques, and an explanation of hardware from the programmer's standpoint- all three things are necessary for writing good (and fast) scientific programs. I know how deep the rabbit hole goes in each of these areas, so I can strongly recommend this book as a good introduction into each for engineers/chemists/physicists that are learning how to write scientific programs. For more about hardware, try the 'write great code' series Write Great Code: Volume 1: Understanding the Machine. For more on numerical analysis, there are many good books, such as Scientific Computing.

This isn't a cookbook for writing the best scientific programs. Sure, there are a few examples, but the underlying principles are what to take away. Despite all my experience, I learned a few tips. I have to say, this book would be better if the authors 1) didn't plug their program at every turn, 2) sat down and narrowed the scope of the book/chapters and critically asked themselves what a given example brings to the discussion, and 3) didn't try to be experts on programming languages. Don't read too deeply into the example implementations. C is my language of choice- I can call them out on many things. I know enough fortran to raise eyebrows at some of the content. The proofreader and technical editor need to be shot- typos and blatant errors are rampant throughout the book (seriously- run spell checker before you hit the "print" button).

"Writing Scientific Software: A guide to good style" is a niche book by Suely Oliveira and David Stewart, who are professors (of computer science and mathematics, respectively) at the University of Iowa. This volume covers a multitude of coding, computer organization, and tool-related issues in connection with scientific programming. It is rounded out by two somewhat extended design examples, one in Fortran and one in C++.

The Good: Oliveira and Stewart provide a nice introduction to the IEEE floating-point standard, as well as wider issues having to do with numerical accuracy and stability. They also include a lot of material on coding and design, using Fortran 90/2003, C, and C++. They show a modicum of Java, though this doesn't seem to serve much of a purpose (they could have included Python instead, see below). Given that they address a community which to a large extent still interacts with Fortran 77, the authors should be commended for advocating fundamental things like avoiding global variables, knowing how to allocate memory dynamically, and using object-oriented as well as generic programming. Similarly, they introduce a number of lessons learned in mainstream software engineering on the importance of good and up-to-date documentation, as well as on the significance of testing. A strong point of the book is a (somewhat scattered) collection of useful tricks. A few examples: on numerical accuracy: add terms in a Taylor series in reverse order (p. 41); on Unix tools: use grep or sed to produce external documentation (p. 81); on debugging: use the __LINE__ macro and also print out the addresses associated with pointers, checking for null values or repeated values, or values that are too close together (pp. 121, 123); on architecture-related issues: avoid if statements inside short inner loops (p. 162). (Of course, the material on architecture and efficiency is far from timeless: even though the book was published in 2006, the authors refer to 64-bit CPUs as to a thing of the future.) Most of the positive aspects of the book are not ground-breaking, but they weren't supposed to be: this is intended as a condensation of knowledge and experience in the areas of numerics, coding, and computer architecture. Covering all that in under 300 pages -- which means that one can read this book cover-to-cover in a short amount of time -- is nothing to sneeze at.

The Bad: the authors provide a list of errata on their website, but these barely scratch the surface. In too many places, the English is awkward (to say the least). A few examples: "Suppose that you are wanting [sic] to solve an ordinary differential equations [sic]" (p. 39); "The most natural way of handling the environment problem goes to object-oriented languages." [?] (p. 103); "Many discrete algorithms do not have efficient algorithms." (p. 152). The uncatalogued errors extend beyond language issues: there are many technical inaccuracies, though it's difficult to tell where the typos end and the true mistakes begin. Here are a few examples, all taken from Part II: the authors fail to distinguish between Fortran subroutines and functions (p. 49); they say that writing *b++ = *a++ is obscure (p. 62), though this is a common idiom in C; they imply that Java's classes predate Fortran's derived data types (p. 90); they think they're overriding a const pure virtual function with a non-const member function (p. 105); they engage in bad programming practice by showing a function defined in a friend declaration (pp. 110-111); they assume that using the inline keyword will inline a function (p. 112), though in reality the keyword is just a request to the compiler; they show a linear combination of C++ and Java and call it C++ (pp. 143-144). On a different note, the authors leave out a number of things: for example, the boost libraries are mentioned only in passing, so there's nothing on multi_array, or on boost (soon to be std::) array, etc. In the whole book, the authors devote one sentence to Python (p. 230). Finally, given that this is a text on scientific software I was surprised by the fact that it contains essentially no mention of multithreading or distributed memory programming.

The people who would have the most to gain by reading this book are probably beginning graduate students, but it is precisely the readers who lack much experience that would have been better served by a more accurate text. (A final example: section 12.8 is called "Basic Linear Algebra Software (BLAS)". By section 16.2, "Software" has morphed into "Subroutines". Both are wrong, however: the "S" in BLAS stands for "Subprograms".) In a nutshell, I think that this volume deserves four stars for its overall choice of material and the authors' generally good advice, and two stars for its slipshod production quality. Overall, three stars.

Alex Gezerlis

Writing Scientific Software: A Guide to Good Style provides college-level audiences studying computer technical writing with a clear guide to writing numerical software. Tips are geared to real-world challenges and scenarios and include plenty of examples, so students learn how to write efficient, workable software, test it for bugs and performance, and more. Examples are provided in Fortran 90, C++ and Java throughout and offer plenty of detail on common scenarios, pitfalls, and the elements of sold scientific software production. A top pick for college libraries.

Diane C. Donovan

California Bookwatch