Hi Mark, Mark H Weaver <m...@netris.org> skribis:
> l...@gnu.org (Ludovic Courtès) writes: > >> As was reported recently by Mark and others, ‘par-map’ would only use >> ncores - 1, because the main thread was stuck in a >> ‘wait-condition-variable’ while touching one of the futures. >> >> The obvious fix is to write ‘par-map’ like this (as can be seen from >> Chapter 2 of Marc Feeley’s PhD thesis): >> >> (define (par-mapper mapper cons) >> (lambda (proc . lists) >> (let loop ((lists lists)) >> (match lists >> (((heads tails ...) ...) >> (let ((tail (future (loop tails))) >> (head (apply proc heads))) >> (cons head (touch tail)))) >> (_ >> '()))))) > > Am I correct in believing that the above code would use the main thread > only for applying 'proc' to the first element of each list? Yes. > In other words, if you have 4 cores and call 'par-map' on a list of 1000 > elements, the main thread will only be used to process 1 out of 1000 > elements, and only 3 cores will be used to process the other 999. > Is that right? No. :-) That’s what would happen with current stable-2.0. But in ‘wip-nested-futures’, the main thread behaves as a worker while waiting for the tail future to complete (which addresses your main concern, I think, as we had discussed on IRC.) Thus, all cores are always used. Thanks, Ludo’.
