rewrite AnalysisBase to be compatible with pmda.parallel.AnalysisBase

[orbeckst]

Expected behavior

MDAnalysis and pmda share code and it is easy to parallelize existing analysis; we don't maintain duplicate code. For instance, I should be able to take a _single_frame() method for MDAnalysis directly in pmda.

Actual behavior

MDAnalysis.analysis.base.AnalysisBase and pmda.parallel.AnalysisBase are different because the serial version can easily share data via the class while the parallel version must have _single_frame() return the results for accumulation. The parallel version has _reduce to handle accumulation in a parallelizable fashion.

Comments

pmda is only one way to parallelize trajectory analysis. However, my understanding is that as soon as one wants to distribute work over processes that cannot communicate by shared memory one needs to change the way in which AnalysisBase works. Therefore, I think that adapting the pmda model is a step in the right direction.

[IAlibay]

Is this still feasible for 2.0?

How is this affected by the proposed changes for mda_cli? Pinging @PicoCentauri here.

[orbeckst]

Not feasible for 2.0

[IAlibay]

@orbeckst shall we move this to 2.1 or do we think it's now become a 3.0 target?

[orbeckst]

Do 3.0 for now and if we can get hands on deck we can move it up.

…

Am 4/6/21 um 12:21 schrieb Irfan Alibay ***@***.***>:  @orbeckst shall we move this to 2.1 or do we think it's now become a 3.0 target? — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or unsubscribe.

[IAlibay]

Looks to me like we're going a different direction here -- closing.

[RMeli]

Linking #4158 for record.