Hidden Serialization

From HPCBugBase

Contents 1 Defect Type Description 2 Entries 3 Advice 4 Symptom 5 Causes 6 Other Findings and Contexts

[edit] Defect Type Description

This defect type is intended to cover "side effects" of parallelization. In typical HPC applications, parallel constructs (parallel library calls, special keywords, directives, etc.) take only a small fraction of the entire code, and the rest is no different from sequential code written in the base language. In a large HPC projects where developers are working as a team, it is not unusual that some team members never write the parallel portion of the code.

Nevertheless, developers who are not directly writing parallel code can still make the classes of defects we are interested in, because serial constructs which were perfectly valid in a sequential contexts can cause correctness and performance defects when they are placed in parallel contexts. Problems tend to aggravate especially when the code is executed in a massively parallel environment.

[edit] Entries

• I/O Hotspots
• Hidden Serialization in Library Functions

[edit] Advice

• Don’t just focus on the parallel code
• Check that the serial code is working on one processor, but remember that the defect may surface only in a parallel context

[edit] Symptom

• Various correctness/performance problems

[edit] Causes

• "Sequential part" tends to be overlooked. Typical parallel programs contain only a few parallel primitives, and the rest of the code is made of a sequential program running in parallel.

[edit] Other Findings and Contexts

Other standard problems with parallelizing the code, such as thread-safety, preemptiveness, etc., can apply to HPC code too. We have not collected samples of this kinds from actual code yet.