form-dev / form / 26817518499

63%

push

github

perf: improve parallelism of SortBotMerge

Currently, the sortbot stage of tform sorting does not achieve good
parallelism. Primarily, this is because the "sortblock" size has
grown with default SmallSize and LargeSize adjustments, such that in
many cases the sortbot levels do not run in parallel at all, because
each thread's output fits in a single block.

This commit adjusts the logic for filling and unlocking the blocks
such that sortbot threads can start work as soon as possible:
 - Only put complete terms into the blocks, no splitting over the blocks.
 - Track the number of terms in each block, and use this when reading
   the data to determine when a block is complete, rather than waiting
   for a term to overlap the "stop" pointer.
 - When filling blocks (in PutToMaster), if a certain (small) minimum
   number of terms has been written, probe if a reading thread is waiting
   on the current block by attempting to lock the previous block. If a
   thread is waiting (so the lock fails), unlock the current block
   immediately and write the term into the next.

Also reduce the MINIMUMNUMBEROFTERMS parameter from 10 to 1, so that
the small+large buffer does not scale (so much) with large MaxTermSize
and many threads, and similarly reduce NUMBEROFBLOCKSINSORT to its minimal
value of 8.

Co-authored-by: AnaIPereira <AnaIPereira@users.noreply.github.com>

60 of 64 new or added lines in 1 file covered. (93.75%)

50887 of 82721 relevant lines covered (61.52%)

5613067.74 hits per line