On Wed, 2015-03-11 at 17:07 +0300, Yuras Shumovich wrote:
On Wed, 2015-03-11 at 08:26 +0100, Kees Bleijenberg wrote:
In the book Parallel and Concurrent Programming in Haskell http://chimera.labs.oreilly.com/books/1230000000929/ch02.html#sec_par-eval-s udoku2 a list of sudokus is solved in parallel.
In version 2 (sudoku2.hs) the program splits the list of sudokus in 2 seperate lists and solves these lists in parallel.
In version3 (sudoku3.hs) parMap is used.
What I don't understand is why in sudoku2 the program has to wait until the parallel computations are finished with rseqs while in sudoku3.hs there is no rseq (not in the main program nor in parMap)? Why can't program sudoku3.hs terminate before all parallel calculations are finished as in Example 2-1. rpar/rpar?
I believe `rseq` is omitted in `sudoku3` for a reason.
In `sudoku2` we spawn two relatively heavy sparks, and the results of them are immediately consumed. So if we don't wait for the sparks, the main thread will start evaluating one of the thunks. Most likely the thunk will be evaluated twice -- by the main thread and the spark. That is called `fizzled` spark.
In `sudoku3` we have a lot of light sparks. When we spawn the last spark, the first one most like is already evaluated (aka `converted`), and its result can be immediately consumed by `length`.
Also I should note that `rseq` seems to be broken now, because it always produces one `converted` and one `fizzled` spark for `sudoku2` in my experience.
Thanks, Yuras
-- Haskell Exists http://blog.haskell-exists.com/yuras