Validation of the Defect Classification Scheme (Jan 06) by Interview

From HPCBugBase

[edit] Background

Following the Validation of the Defect Classification Scheme (Jan 06) by Survey, the second validation of the Defect Classification Scheme (Jan 06) was performed at the level of individual defects as well as the classification level. We interviewed a professor who has taught an HPC course for several years, along with his teaching assistant who has been involved in multiple iterations of the course. The professor has many years of experience in HPC development. The interview was conducted over the phone. Before the interview, we prepared presentation material containing concrete examples of defects categorized by type. During the interview, we went through each defect example and examined if it looked reasonable.

[edit] Interview Results

The interviewees agreed that all defect examples were observed quite often. They also agreed the classification scheme was reasonably defined, but they commented about wording as follows:

• The defect type Serial Constructs is confusing. A suggested name is Side-effect of Parallelization.
• The defect type Load Balancing is too specific, as some defects categorized as this type are addressing different kind of performance problem (scheduling). A suggestion is to rename it to Performance to accommodate both kinds of problems.

Finally, they pointed out the defect which had not been identified in the previous analysis. In the defect with the use of a pseudo-random sequence, the implementation of the rand() function causes hidden serialization. It leads to a performance problem when this function is called by many processors simultaneously.