Chunk distribution strategies

[franzpoeschel]

Based on topic-available-chunks, see #802.
Based on #1043

Adds strategies for guiding applications to efficient parallel chunk loading patterns.
Idea: PR #802 lets reading applications inquire the chunks that are available for loading in the backend. This PR adds a function chunk_assignment::assignChunks, taking such a ChunkTable as well as a strategy as input and produces a further ChunkTable that may be used as a load pattern.

Implemented strategies:

• Round Robin: Assign chunks to reading processes in turn.
• by hostname: Assign chunks to processes on the same host. Parameterized by the interior distribution strategy.
• by cuboid slice: Assign chunks according to a slicing into n_ranks hyperslabs of the data space.
• by binpacking: Assign chunks in an approximatively balanced manner via an approximation algorithm for the NP complete binpacking problem.

Also, add an attribute to the global Series object (currently called rankMetaInfo, I'm sure we'll change this) that can be used to write string-formatted meta information for the single writing processes and assign meaning to them (e.g. hostnames). Used by the hostname-based strategy.

TODO:

• Python bindings
• Testing, mostly done manually so far
• Cleanup
• Documentation
• This PR adds many MPI-based methods to the Series class. It would probably be good to first separate MPI and non-MPI headers, and then to add this functionality to the MPI headers. Still a goal, but no longer relevant for this particular PR.

[lgtm-com]

This pull request introduces 1 alert when merging 96841f9 into b2664b7 - view on LGTM.com

new alerts:

• 1 for Catching by value

[lgtm-com]

This pull request introduces 1 alert when merging 6da1db3 into a6287fc - view on LGTM.com

new alerts:

• 1 for Catching by value

[lgtm-com]

This pull request introduces 1 alert when merging 43b1e31 into f6b0054 - view on LGTM.com

new alerts:

• 1 for Catching by value

[lgtm-com]

This pull request introduces 1 alert when merging 5f2b347 into f6b0054 - view on LGTM.com

new alerts:

• 1 for Catching by value

[lgtm-com]

This pull request introduces 1 alert when merging bb7fa45 into 9b349eb - view on LGTM.com

new alerts:

• 1 for Catching by value

[franzpoeschel]

@ax3l ping

[lgtm-com]

This pull request introduces 1 alert when merging ebea6cf into 86a15ba - view on LGTM.com

new alerts:

• 1 for Unused local variable

[franzpoeschel]

I've added Python bindings now, including usage of one distribution strategy in openpmd-pipe as an integrated usage example.

[franzpoeschel]

One issue of this approach is that the attribute rankMetaInfo is a vector of as many strings as there are parallel writers which I suspect is part of the reason why we saw scaling issues in some benchmarks.
Instead, a parallel dataset should preferrably be written.

EDIT: I've transformed this to a parallel dataset now. Needs some more checking for char portability. Also, we should probably buffer these values on reading, since the table can only be read in steps in which it was written. In short: Make this less like an attribute that can be set and inquired, and rather more automatic during Series construction.

[ax3l]

Can you instead comment in the line 1335 pls?

[ax3l]

Nice!

Please add the small CMake update and then it can go in.