feedback needed on explaining Amdahl's law

[psteinb]

I am closing in on reaching version 1.0 for hpc-in-a-day. After this is done, I'll start sending PRs to this repo.

After some debate, I decided to include Amdahl's law in my material. I know that hpc-novice is likely not to adapt this part of my course about parallelisation. But in any case, I'd love to get some feedback on whether I got the level right:
https://psteinb.github.io/hpc-in-a-day/03-02-parallel-estimate/

I'll close this issue once the first comments come in on hpc-in-a-day or when my next workshop is over on Jan 24th.
Thanks in advance - P

[shwina]

@psteinb

Overall, I think it's absolutely great. I especially like the attention to detail in both your explanation and in your figures.

However, I feel that using Amdahl's law to compare the speedup from two parallelization strategies for the same program may not be as instructive as looking at two different programs, one that is mostly parallel, and one that is mostly serial (e.g., does lots of serial I/O).

I think Amdahl's Law can be better used to drive home the point that sometimes parallelization is not worth it, rather than to compare two ways to parallelize the same code - does that make sense?

[psteinb]

Excellent comment - I think you are referring to the other side of the same medal which I want to bring across. I will come up with an exercise targeting this use of the law.

[markcmiller86]

FWIW, I think questions of parallelization strategies and speedups are well beyond the kind of a material an HPC novice will require especially if they are simply looking for how to login and submit some pre-installed application there to the batch system.