I replaced my blocking queue sitting between pipeline actors (threaded consumer/producers) and speed of my word freq program dropped from 5.2s to 3.5s on 5M file simply by replacing the blocking queue with my new ping-pong buffered version. That is close to the 3.0s achieved by the nonthreaded version.
// threaded, pipeline version; 3.5s (from 5.2s) on 5M file
f => Words() => { string w | ... }
// nonthreaded, nested map operation on big list; 3.0s
f.lines() => a;
a:{ string line | line.split():{ string w | ... }};
Buffer size is 1000. When I drop to 100, slows down by .2s. When I increase to 4000, no change. 40000, slows down by .5s. Size 400 seems to be slightly faster. Default queue size set to 400 now.
Thanks to Sriram Srinivasan for the idea of ping-ponging the buffers.