Distributed cluster gather/sort/filter/match operations

[basaks]

By # 41c83f7, we can process a small number of events, say upto, 5k events. This works on single process and in memory sort/filter/joins using pandas dataframe.

However, we need to process upwards of 500k+ events. Just from ISC and Engdahl we have 300k+ events.

[basaks]

@alexgorb We can test 3d inversion functionality using a small number of events ( a few thousand events). Should we finish this parallelisation now before moving onto other tasks?

[basaks]

corresponding jira: https://gajira.atlassian.net/browse/PST-227

[basaks]

1. So far, gather of arrivals is optimally distributed.
2. Distributed median being difficult to compute, we still have a single process sort and median computation. Improved performance by using pandas throughout and avoiding for loop. For details see Improve median filter performance during 3d travel time inversion input generation #49.
3. Matching is still single process and very efficient for the size of our data by the matching stage.

Closing.