Filter the sources in parallel, so that we can remove source_max_weight parameter

[Michele Simionato]

There is some cost here, since we will have to perform the filtering twice. Some performance test is in order.

[Michele Simionato (michele-simionato)]

My analysis in real life use cases (especially the USA computations) shows that the duplication is negligible. For instance you may have 200,000 sources, 180,000 are filtered out, and therefore the duplicate filtering happens only for 20,000 sources, thus taking a negligible amount of time. On the other hand, the speedup is minor because a lot of the time is spend in transferring the sources back and forth. Still, it is worth doing because we can remove the parameter source_max_weight and keep only the parameter concurrent_tasks which is much more transparent to the end user.

[Michele Simionato (michele-simionato)]

Here are some numbers for the USA. Filtering/splitting in the master node (1 core) takes 1,493s, on workers (256 cores) takes
429s , i.e. the speedup is only of a factor of 3.5.

[Michele Simionato (michele-simionato)]

With the full model (153,000+ sites) the `process_sources` time improves from 9361s to 1130s (8x speedup).

[Michele Simionato (michele-simionato)]

https://github.com/gem/oq-engine/pull/1494