Backend NML Parser: Performance Improvements

[fm3]

I had wrongly assumed the bulk of the upload nml complexity to be in validateTrees.

Analysis with a 350 MB real-world NML yielded:

45 s  load xml from input stream
961 s  parse trees
3 s  splitMultiComponentTrees
6 s  validateTrees
0.7 s  send to tracingstore & save

This PR tackles two properties of parseTrees:
(1) the accessing of XMLNode attributes via shortcut method \ is not very efficient, new helper method getSingleAttribute (speedup factor ×2–4)
(2) replace inefficient handling of comments and branchpoints by hashmap lookup (speeup factor ×10–15 for this particular NML, which notably has 80k node comments)

so the parse trees line after this change reads with speedup factor ×36

27 s  parse trees

with the others unchanged.

Other changes:

• increase request timeouts from 1000s (16.6min) to 10000s (166min)
• use the same request timeouts in dev mode (has to be defined in built.sbt)
• when killing the proxy.js, send sigkill rather than sigterm to child processes, avoiding port-in-use problems

TODO

• use getSingleAttribute in remaining places (won’t bring much more speedup, just for consistency)
• cleanup (remove time-logging, re-enable startup scripts etc)

Steps to test:

• upload very large NML via dashboard
• should be faster, should yield the same SkeletonTracing

---

• Updated (unreleased) changelog
• Ready for review

[youri-k]

LGTM