Spread-Filling Placement

[mvousden]

This changeset:

• Introduces a new algorithm (spread), which spreads devices across the engine to minimise device density as much as possible.
• Adds a small, hackish testing framework for placement, and adds a couple of tests to demonstrate the new algorithm.
• Adds a new GraphI_t method, DevicesByType, which computes the total for each normal device type in the graph instance.
• Modifies Placer::redistribute_devices_in_core to respect both global and graph-instance-specific threads-per-core constraints.

The placement algorithm has been tested using the new testing framework. The new testing framework has also been tested with "impossible" scenarios created by constraining devices-per-thread.

See also the documentation changes here (https://github.com/POETSII/orchestrator-documentation/pull/11), which I will merge when both PRs are approved.

[mvousden]

One other thing that threw me was the placement dump - I went looking in the microlog for the dump info but there was nothing there.

I also found this, and included it in my documentation changeset for #186 (https://github.com/POETSII/orchestrator-documentation/pull/12/files)

[heliosfa]

One other thing that threw me was the placement dump - I went looking in the microlog for the dump info but there was nothing there.

I also found this, and included it in my documentation changeset for #186 (https://github.com/POETSII/orchestrator-documentation/pull/12/files)

Ah, found it buried in there. It might be nice (but not a show stopper) to mention it in the microlog as well, e.g. just a "dumped placement data to X" line. Empty micrologs are a little bit annoying...

Not a showstopper though.