How To Do Parallelism Without Getting Egg On Your Face
The other day I put together a trivial parallel code example that was going to be used in a 'how-to' document. The code was nothing much more complicated than a parallel 'hello world'. The sort of thing you could write while asleep.
I then sent the code to three colleagues for review. Well, embarrassingly the code came back, the reviewers pointing out that the code had a data race and the results of the program were inconsistent.
What had I done wrong? Simple, because the code was 'easy', I had fallen into the trap of thinking that I could spot every problem by reading through the code. Now, if I'd have been sensible and used tools that come with Intel Parallel Studio then I would not now have egg on my face.
When adding parallelism to a piece of code, a well tested strategy is to do it in several stages -- Analysis, Implementation, Verification, and Tuning. Typically these stages are repeated as parallelism is incrementally added to the code. Intel Parallel Studio has tools to help in each of these phases.
Additionally (a reminder to myself!), when changing a program from serial to parallel, you should really make sure the following questions are all addressed.
Questions to ask at the Analysis phase
- Is my program parallel?
- Where is the best place to parallelise my program?
- How can I get my program to run faster?
- What's the expected speedup?
Questions to ask at the Implementation phase
- How difficult?
- Is my code still working?
Questions to ask at the Verification phase
- Is the parallelism correct?
- Do I have deadlocks or data races?
- Do I have memory errors?
- Does my program still work as intended?
Questions to ask at the Tuning phase
- Do my tasks do equal amounts of work?
- Is my application scalable?
- Is the threading running efficiently?
If only I had used Parallel Inspector at the Verification stage, I would now not look quite so stupid.
Let's hope only three of my colleagues spotted the error of my ways. I think this is a case of Physician heal yourself.