From: Damien Mattei <damien.mattei@gmail.com>
To: guile-user <guile-user@gnu.org>, guile-devel <guile-devel@gnu.org>
Subject: Re: map-par slower than map
Date: Thu, 13 Oct 2022 10:20:19 +0200 [thread overview]
Message-ID: <CADEOaddYEeBvYsXt=0F+AsX_7BhhEOAG=t7AUhRgmMH2Q-5j4Q@mail.gmail.com> (raw)
In-Reply-To: <CADEOadcxSDvWEd7fQMWUAPQXejNU8WV_AU9OV7fh+qu_Xt_1yQ@mail.gmail.com>
ok 'parallelization code and hash table' in google show no real solution, i
have another idea,instead of function-unify-two-minterms-and-tag ,just
unify the minterms and tag them (which use hash table) later out of the //
region, after if it still sucks i will drop list for vectors, making it
step by step will show where is the bottleneck, let me just a few time and
i will go back to the mailing lists with interesting results i hope....
On Thu, Oct 13, 2022 at 9:40 AM Damien Mattei <damien.mattei@gmail.com>
wrote:
> ok , i think the problem comes both from my code and from guile parmap so.
> Obviously parmap could be slower on other codes because of the nature of
> list i think, it is hard to split a list in sublist and send them to thread
> and after redo a single list, i better use vector.
> As mentioned and forget by me, i apologize, i use an hash table which is a
> global variable and mutex can be set on it , no dead lock , here but it
> slow down the code than it is dead for speed , but not locked.
>
> The examples given are good but i do not want to write long and specific
> code for //, for // must a ssimple as OpenMP directives (on arrays) not be
> a pain in the ass like things i already did at work with GPUs in C or
> Fortran,that's nightmare and i do not want to have this in functional
> programming.
>
> ok i will try to modify the code but i do not think there is easy solution
> to replace the hash table, any structure i would use would be global and
> the problem (this function is not pure, it does a side-effect),at the end i
> do not think this code could be // but i'm not a specialist of //. I think
> this is a parallelization problem, how can we deal with a global variable
> such as an hash table?
>
> On Wed, Oct 12, 2022 at 11:55 PM Olivier Dion <olivier.dion@polymtl.ca>
> wrote:
>
>> On Wed, 12 Oct 2022, Damien Mattei <damien.mattei@gmail.com> wrote:
>> > Hello,
>> > all is in the title, i test on a approximately 30000 element list , i
>> got
>> > 9s with map and 3min 30s with par-map on exactly the same piece of
>> > code!?
>>
>> I can only speculate here. But trying with a very simple example here:
>> --8<---------------cut here---------------start------------->8---
>> (use-modules (statprof))
>> (statprof (lambda () (par-map 1+ (iota 300000))))
>> --8<---------------cut here---------------end--------------->8---
>>
>> Performance are terrible. I don't know how par-map is implemented, but
>> if it does 1 element static scheduling -- which it probably does because
>> you pass a linked list and not a vector -- then yeah you can assure that
>> thing will be very slow.
>>
>> You're probably better off with dynamic scheduling with vectors. Here's
>> a quick snippet I made for static scheduling but with vectors. Feel
>> free to roll your own.
>>
>> --8<---------------cut here---------------start------------->8---
>> (use-modules
>> (srfi srfi-1)
>> (ice-9 threads))
>>
>> (define* (par-map-vector proc input
>> #:optional
>> (max-thread (current-processor-count)))
>>
>> (let* ((block-size (quotient (vector-length input) max-thread))
>> (rest (remainder (vector-length input) max-thread))
>> (output (make-vector (vector-length input) #f)))
>> (when (not (zero? block-size))
>> (let ((mtx (make-mutex))
>> (cnd (make-condition-variable))
>> (n 0))
>> (fold
>> (lambda (scale output)
>> (begin-thread
>> (let lp ((i 0))
>> (when (< i block-size)
>> (let ((i (+ i (* scale block-size))))
>> (vector-set! output i (proc (vector-ref input i))))
>> (lp (1+ i))))
>> (with-mutex mtx
>> (set! n (1+ n))
>> (signal-condition-variable cnd)))
>> output)
>> output
>> (iota max-thread))
>> (with-mutex mtx
>> (while (not (< n max-thread))
>> (wait-condition-variable cnd mtx))))
>> (let ((base (- (vector-length input) rest)))
>> (let lp ((i 0))
>> (when (< i rest)
>> (let ((i (+ i base)))
>> (vector-set! output i (proc (vector-ref input i))))
>> (lp (1+ i)))))
>> output))
>> --8<---------------cut here---------------end--------------->8---
>>
>> --
>> Olivier Dion
>> oldiob.dev
>>
>
next prev parent reply other threads:[~2022-10-13 8:20 UTC|newest]
Thread overview: 27+ messages / expand[flat|nested] mbox.gz Atom feed top
2022-10-12 17:19 map-par slower than map Damien Mattei
2022-10-12 18:45 ` Maxime Devos
2022-10-12 20:20 ` Damien Mattei
2022-10-12 20:27 ` Damien Mattei
2022-10-12 21:29 ` Zelphir Kaltstahl
2022-10-14 8:21 ` Damien Mattei
2022-10-14 8:38 ` Zelphir Kaltstahl
2022-10-17 13:17 ` Damien Mattei
2022-10-22 16:01 ` Damien Mattei
2022-10-23 1:06 ` Damien Mattei
2022-10-25 9:07 ` Mikael Djurfeldt
2022-10-25 9:11 ` Mikael Djurfeldt
2022-10-25 14:09 ` Damien Mattei
2022-10-12 21:44 ` Maxime Devos
2022-10-12 21:55 ` Olivier Dion via Developers list for Guile, the GNU extensibility library
2022-10-13 7:40 ` Damien Mattei
2022-10-13 8:20 ` Damien Mattei [this message]
2022-10-13 9:10 ` Olivier Dion via Developers list for Guile, the GNU extensibility library
2022-10-13 10:44 ` Damien Mattei
2022-10-13 11:00 ` Olivier Dion via Developers list for Guile, the GNU extensibility library
[not found] ` <CADEOadfovi8s3OxRcssWOuOW8jjHoL9Z7pD_5FstSm=ZkBHP8g@mail.gmail.com>
2022-10-13 11:57 ` Fwd: " Damien Mattei
2022-10-13 12:36 ` Damien Mattei
2022-10-13 12:41 ` Olivier Dion via Developers list for Guile, the GNU extensibility library
2022-10-13 13:43 ` Damien Mattei
2022-10-13 14:06 ` Olivier Dion via Developers list for Guile, the GNU extensibility library
2022-10-13 14:10 ` Damien Mattei
2022-10-13 14:21 ` Damien Mattei
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
List information: https://www.gnu.org/software/guile/
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to='CADEOaddYEeBvYsXt=0F+AsX_7BhhEOAG=t7AUhRgmMH2Q-5j4Q@mail.gmail.com' \
--to=damien.mattei@gmail.com \
--cc=guile-devel@gnu.org \
--cc=guile-user@gnu.org \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).