Thanks Rodolfo, I did some further experiments which confirm your results. I noticed that results vary quite a bit run to run so I ran several times from clean and took the median:
prime-pi/acc 4000000 with original make-coroutine: cpu time: 2808 real time: 2879 gc time: 1000 prime-pi/acc 4000000 w/ stream-ref precall with original make-coroutine: cpu time: 3214 real time: 3245 gc time: 1342 prime-pi/acc 4000000 with alternate make-coroutine: cpu time: 2886 real time: 2996 gc time: 856 prime-pi/acc 4000000 w/ stream-ref precall and with alternate make-coroutine: cpu time: 2901 real time: 2968 gc time: 1078 prime-pi 4000000 w/ stream-ref precall and with original make-coroutine: cpu time: 5054 real time: 5063 gc time: 2810 It is still interesting that prime-pi performed so badly with the alternate coroutine (and without the precall cludge). I reran it this way and had to break the process when it ran over 5 minutes. BTW, with the cludge and the alternate make-coroutine it ran fine, yesterday's results for that case are bogus as I believe that after a bad run you need to actually bring down DrRacket (not just hit the run button) in order to clean up memory. I think we also confirm that you have a better machine than I :-) -joe On Wed, Feb 22, 2012 at 8:40 AM, Rodolfo Carvalho <rhcarva...@gmail.com>wrote: > Hello Joe, > > > 2012/2/22 Joe Gilray <jgil...@gmail.com> > >> Hi Matthew, >> >> Thanks for your help. I really can't even pretend to understand the >> make-coroutine code but this is what I'm seeing: >> >> Original make-coroutine, without stream-ref pre-call => out of memory >> (1GB) when running (prime-pi 500000) >> > > > I tried to run your code here and I also get out of memory in the first > case. > > > > >> Original make-coroutine, with stream-ref pre-call => runs (prime-pi >> 500000) in about .4 sec >> > > > Just as a reference to the other results, here it gives: cpu time: 328 > real time: 328 gc time: 94. > > > > >> Alternate make-coroutine, with or without stream-ref pre-call => runs >> (prime-pi 500000) in about 28 sec >> >> > With stream-ref pre-call: cpu time: 343 real time: 344 gc time: 78 > Without: cpu time: 19609 real time: 19719 gc time: 9172 > > > > > I changed `prime-pi' into `prime-pi/acc' as below and then I got the best > results: > > (define (prime-pi/acc n) > (let loop-count ([candidate-stream primes] > [sum 0]) > > (if (> (stream-first candidate-stream) n) > sum > (loop-count (stream-rest candidate-stream) (add1 sum))))) > > > With original make-coroutine: cpu time: 234 real time: 235 gc time: 0 > With Matthew's make-coroutine: cpu time: 187 real time: 187 gc time: 0 > > > > All measurements made from within DrRacket using (time (prime-pi 500000)) > or (time (prime-pi/acc 500000)). > > []'s > > Rodolfo >
____________________ Racket Users list: http://lists.racket-lang.org/users
