Re: map-par slower than map

[Damien Mattei]

Hello Zelphir,
i should had written first about the mathematical problem first but i did
not want to go into other than computer things. But i expose here the main
idea briefly: i try to solve a mathematical conjecture (
https://en.wikipedia.org/wiki/Conjecture ) using logic expressions but not
in the common sense of Boole's algebra, at some point i shift to
Probability logic and i need to simplify, minimalize logic expressions in
disjunctive form, this is my idea. I will give a few link for the curious
and i hope to publish something in the next year. At the beginning it is a
computer problem well know in logic: the minterms of the hash table i use
are described here:
https://en.wikipedia.org/wiki/Canonical_normal_form#Minterms
and the algorithms are here:
https://en.wikipedia.org/wiki/Quine%E2%80%93McCluskey_algorithm
https://en.wikipedia.org/wiki/Petrick%27s_method
The minterms come from a set product :
https://en.wikipedia.org/wiki/Cartesian_product
(space complexity is polynomial)
 but the whole logic problem is an NP-hard problem so i do not expect it to
be fast for more than 10 variables, it works well on one processor, and i'm
now checking it on multiple core. I could have used Mathematica or Python
sympy but i choosed many years ago to build a system in Scheme because
there is no free software, Maxima does not support well logic (there was a
no more supported module for Zhegalkin polynomials
https://en.wikipedia.org/wiki/Zhegalkin_polynomial ) i did not want to use
a commercial product such as Mathematica and i did not know Python Sympy
ten years ago... but now i will use it to cross-check my  Scheme results. I
hope to do that this week, if it is ok i will not modify the Scheme code
any more.
The hash table was the more obvious structure to use due to the nature of
algoritms here, i can not use arrays.
I did not expected the compiler to solve the concurrent access to the hash
table, i know i was going into problems, and i solve the problem with
arrays in the // region and put back data after in the hash table in a non
// region.It is okay now.

For the one interested (i apologize because this is out of subject in the
Guile mailing list) here is a few interesting links about Probability and
Logic:
Theodore Hailperin wrote the best articles in my opinion:

https://projecteuclid.org/journals/notre-dame-journal-of-formal-logic/volume-25/issue-3/Probability-logic/10.1305/ndjfl/1093870625.full

https://plato.stanford.edu/entries/logic-probability/

https://www.amazon.fr/Booles-logic-probability-contemporary-foundations/dp/0444110372

Best regards,
Damien


On Mon, Oct 24, 2022 at 1:18 AM Zelphir Kaltstahl <
zelphirkaltstahl@posteo.de> wrote:

> Hello Damien!
> On 10/23/22 03:06, Damien Mattei wrote:
>
> after intense coding i finally got the good results,
> my assumption about the global variable hash table was true ,it is
> incompatible with 'future : the competition for the writing into the hash
> table breaks the code.
>
> If i do writing in hash table out of // region all is ok:
>
> a simple function to extract the hash table access:
>
> (define (tag-minterms i umt)
>   (when umt
>      {mt1 <+ (first {minterms-vector[i]})}
>      {mt2 <+ (second {minterms-vector[i]})}
>      {minterms-ht[mt1] <- #t}
>      {minterms-ht[mt2] <- #t}))
> [...]
>
> I am not sure what exactly the problem is, which you are trying to
> calculate, but it looks fairly mathematical to me. Maybe you do not need to
> share state globally at all? Maybe you can find a way to avoid global
> shared state? I am guessing, that you want to avoid duplicated computation
> of part terms?
>
> Of course,if you have global state and do not have a synchronization
> construct (!) for accessing the hash table, I would expect things to go
> wrong at some point, with non-reproducible results. I do not think that
> futures are to blame here, or parallel map in that case.
>
> With a synchronization construct, some kind of mutex, your bottle neck
> might just become that mutex. Up to you to measure that ; )
>
> Would be nice, if you found a clever way to separate the problems into
> disjunct parts and then solve them without global state.
>
> Regards,
> Zelphir
>
> --
> repositories: https://notabug.org/ZelphirKaltstahl
>
>