I can bet that some 2 min 50 seconds in baseline took by byte -> String conversion. Java is pretty bad at this, especially if default Charset is UTF8. Even using ASCII at Files.lines() call may improve performance dramatically.
All leaders work directly with bytes. All other tricks is just for sub-second improvements.
Did solve similar problem with log files.
Yeah strings are very expensive, saw huge gains in my submission after removing them. Parallelization gets the runtime down to 30s , After removing strings you get to around 15s on the eval machine, the rest of the tricks are still very relevant to get to sub 10s.
