Customer Reviews

Parallel Computing in Quantum Chemistry

5 star:		(1)
4 star:		(0)
3 star:		(0)
2 star:		(1)
1 star:		(0)

 

 

This is a well-written book aimed at researchers in the field of quantum chemistry -- from graduate students to long-standing experts -- who require a concise and clear description of the most important problems facing efforts to parallelize ab initio quantum chemical programs. Given the rapid emergence of new petascale computer systems containing thousands to even millions of computing cores, the timing of this book is fitting.

Full disclosure: I know the authors of this book well. I have published two peer-reviewed journal articles with Dr. Janssen and one with Dr. Nielsen, and I received lab-directed research and development funding from Sandia National Labs through a Department of Energy project of Dr. Janssen's in 2001-2004. In addition, I reviewed the proposal for the book for Taylor and Francis publishers. However, I was not involved in the writing of the book at all. I purchased it of my own accord, and I am writing this review only because I am very impressed with the finished product.

I found the book to be tremendously enlightening. In the first half, the authors provide an overview of essential aspects and tools of parallel computing: hardware, network topology, message-passing software and methods, threading, load-balancing, etc. In addition, they give a fairly detailed explanation of methods for modeling the parallel performance (speedup and efficiency) of algorithms, as well as aspects of parallel program design. One of the strengths of the book is the way the authors make their points clearer by constantly returning to a few specific examples, including matrix-vector multiplication and the second-order Moller-Plesset perturbation theory (MP2) algorithm.

They then make use of the fundamentals developed in the first half of the book to address several key problems in quantum chemical programs: two-electron repulsion integral evaluation, the integral-direct Hartree-Fock method, as well as canonical and local MP2 energy calculations. These provide fertile soil for discussions of load balancing, collective versus one-sided communication, and hybrid (simultaneous shared- and distributed-memory) parallel methods. Each example is well-supported by performance models that provide a clear analysis of the scalability of each algorithm.

My only criticism of the book is that it stops too soon. The numerous problems associated with parallel implementation of more advanced and complicated methods, especially coupled cluster theory, are not discussed, and I would have enjoyed reading the authors' take on this area of on-going research.

Nevertheless, I believe this book will prove to be extremely valuable to those developing quantum chemical program for emerging massively parallel supercomputers. The authors' perspective on the parallelism problem is state-of-the-art, and our field would be wise to listen carefully to what they have to say.

First let me say I'll agree with crawdad's statement "it stops too soon", but we differ on how. The book takes a far too cursory description of the parallelizing quantum chemistry codes.

The first 2/3rds of the book are pretty basic parallel computing calculations. The authors tie it to some examples. But there seems to be a disconnect from the topic at the front of the chapter and the examples at the end. The chapter on computer/network architectures is a good example; after referring to seven different network topologies in abstract terms, only two real world examples are given, and that 10 pages later.

In Chapter 5 the authors spend a lot of time working through Amdal's & Gustafson's laws which provide basic models of computation time, but then they go unused for a more precise big O() notation. And Amdal & Gustafson are never heard from again.

One quirk of the way the formulas are written needs to be corrected too. Many of the communications formulas scale logarithmically by processor counts, but the formulas are not formatted to distinguish this. For example, log2pa is actually (log2(p))*a. It reads as though the a (latency) is a part of the log, but it is not.

When we get around the second section, I have a real hard time following. Realize that I come from a Computer Science background not from Chemistry/Physics and simultaneously was working through Lowe's Quantum Chemistry. So maybe it would make sense to a more seasoned chemist. But I have a hard time figuring out precisely what the author is talking about. The code samples are not all that much help. They are in a very high-level pseudo code. So figuring out which element is which and what data is needed where through code inspection is not an option.

I don't want to be entirely critical here. It is very readable despite the problems noted above. There is a lot of good information in the book, but it needs to be much more in depth.