Computational Scientists Still Don't Get It
A workshop called “Software Issues in Computational Science and Engineering” is running in Uppsala, Sweden, this August. Here’s their blurb:
Software for numerical computations faces multiple challenges. The software should be easy to use. Ideally, adaptation to new applications should be flexible, and extension to incorporate new numerical techniques straightforward. At the same time the software should execute extremely efficiently on various high-performance computing platforms. Accuracy and robustness are other key features. The overall challenge is to find ways to construct numerical software so that all these different goals are met simultaneously.
Once again, there’s no mention of making the programs right—nothing about testing, nothing about tracking results so that when a bug does appear you know what you should retract, nothing. I’m sure the organizers would say, “Oh, that’s part of accuracy,” but I’ve been part of enough discussions to know that when numerical scientists say “accuracy and robustness”, they’re talking about algorithms, not about coding bugs. Given stories like this one, it’s a revealing oversight.