[erlang-questions 51] Re: A sudoku solver in Erlang compared to Python

Evans, Matthew <>
Sun Mar 27 17:49:12 CEST 2011


Good call on using atoms instead of lists. I wouldn't have imagined it would be that much faster. So on my computer the new Erlang implementation is much faster than python:

Erlang with parse transforms and HiPE:

[ ~]$ ./sudoku
expanding...
expanding...
expanding...
expanding...
Solved 50 of 50 puzzles from easy50.txt in 0.059767 secs (836.58 Hz)
(92538 total eliminations, avg 1850.76, median 1811, max 2628, min 1767).
Solved 95 of 95 puzzles from top95.txt in 0.344457 secs (275.80 Hz)
(901201 total eliminations, avg 9486.33, median 6267, max 56820, min 1792).
Solved 11 of 11 puzzles from hardest.txt in 0.025747 secs (427.23 Hz)
(33653 total eliminations, avg 3059.36, median 3023, max 5346, min 1786).


Python:

[ ~]$ ./sudoku.py
All tests pass.
Solved 50 of 50 puzzles from easy50.txt in 0.530000 secs (94.34 Hz)
 (33059 total eliminations, avg 661.00, median 648, max 830, min 648).
Solved 95 of 95 puzzles from top95.txt in 3.980000 secs (23.87 Hz)
 (221997 total eliminations, avg 2336.00, median 1492, max 11512, min 648).
Solved 11 of 11 puzzles from hardest.txt in 0.150000 secs (73.33 Hz)
 (9436 total eliminations, avg 857.00, median 817, max 1198, min 648).
[ ~]$


The initial implementation in Erlang:

[ ~]$ ./sudoku
./sudoku:4: Warning: variable 'Args' is unused
Solved 50 of 50 puzzles from easy50.txt in 0.492551 secs (101.51 Hz)
(92538 total eliminations, avg 1850.76, median 1811, max 2628, min 1767).
Solved 95 of 95 puzzles from top95.txt in 3.120420 secs (30.44 Hz)
(901201 total eliminations, avg 9486.33, median 6267, max 56820, min 1792).
Solved 11 of 11 puzzles from hardest.txt in 0.147977 secs (74.34 Hz)
(33653 total eliminations, avg 3059.36, median 3023, max 5346, min 1786).

________________________________________
From: Kresten Krab Thorup []
Sent: Sunday, March 27, 2011 9:28 AM
To: Evans, Matthew
Cc: Ahmed Omar; ; Andreas Pauley
Subject: Re: [erlang-questions 41] Re: A sudoku solver in Erlang compared to Python

For reference, I tried to run this with Erjang and R14B01 (with BEAN and HiPE/o3) on my machine ...

Oh, and I changed the square names to be atoms in stead of [X,Y] pairs, which speeds up things quite a lot too ... e.g. :

squares() ->
   %% Returns a list of 81 square names, including "A1" etc.
    [erlang:list_to_atom([X,Y]) || X <- "ABCDEFGHI", Y <- "123456789"].


BEAM
> sudoku_main:run_solver(["solve"]).
Solved 50 of 50 puzzles from easy50.txt in 0.678921 secs (73.65 Hz)
  (92538 total eliminations, avg 1850.76, median 1811, max 2628, min 1767).
Solved 95 of 95 puzzles from top95.txt in 4.373277 secs (21.72 Hz)
  (901201 total eliminations, avg 9486.33, median 6267, max 56820, min 1792).
Solved 11 of 11 puzzles from hardest.txt in 0.198780 secs (55.34 Hz)
  (33653 total eliminations, avg 3059.36, median 3023, max 5346, min 1786).
ok

Erjang
> sudoku_main:run_solver(["solve"]).
Solved 50 of 50 puzzles from easy50.txt in 0.653891 secs (76.47 Hz)
  (92538 total eliminations, avg 1850.76, median 1811, max 2628, min 1767).
Solved 95 of 95 puzzles from top95.txt in 4.354102 secs (21.82 Hz)
  (901201 total eliminations, avg 9486.33, median 6267, max 56820, min 1792).
Solved 11 of 11 puzzles from hardest.txt in 0.208266 secs (52.82 Hz)
  (33653 total eliminations, avg 3059.36, median 3023, max 5346, min 1786).
ok

r14b01 / HiPE / o3
> sudoku_main:run_solver(["solve"]).
Solved 50 of 50 puzzles from easy50.txt in 0.611918 secs (81.71 Hz)
  (92538 total eliminations, avg 1850.76, median 1811, max 2628, min 1767).
Solved 95 of 95 puzzles from top95.txt in 3.759281 secs (25.27 Hz)
  (901201 total eliminations, avg 9486.33, median 6267, max 56820, min 1792).
Solved 11 of 11 puzzles from hardest.txt in 0.169039 secs (65.07 Hz)
  (33653 total eliminations, avg 3059.36, median 3023, max 5346, min 1786).
ok

This version does not use pmap (Erjang's scheduler still needs some work), but does use atoms for square names, and ct_expand.

Cheers,

Kresten


On Mar 26, 2011, at 23:33 , Evans, Matthew wrote:

> Awesome point, I was going to actually suggest that since this is all calculating static data he just pre-calculates it and puts it in a -define. Or even better, if he is brave he could write a parse transform.
>
> Using Ulf's ct_expand module (http://forum.trapexit.org/viewtopic.php?p=20260) I get even better performance:
>
> Without the parse transform (but using native):
>
> 4> sudoku:print_results("top95.txt", "\n").
> Solved 95 of 95 puzzles from top95.txt in 0.827966 secs (114.74 Hz)
> (901201 total eliminations, avg 9486.33, median 6267, max 56820, min 1792).
> ok
>
>
> With a parse transform:
>
> 2>  sudoku:print_results("top95.txt", "\n").
> Solved 95 of 95 puzzles from top95.txt in 0.469908 secs (202.17 Hz)
> (901201 total eliminations, avg 9486.33, median 6267, max 56820, min 1792).
>
>
> The code looks like:
>
> -compile({parse_transform, ct_expand}).
>
>
> squares() ->
>    %% Returns a list of 81 square names, including "A1" etc.
>    ct_expand:term(
>        [[X,Y] || X <- "ABCDEFGHI", Y <- "123456789"]
>    ).
> col_squares() ->
>    %% All the square names for each column.
>    ct_expand:term(
>        [[[X,Y] || X <- "ABCDEFGHI", Y <- [C]] || C <- "123456789"]
>      ).
> row_squares() ->
>    %% All the square names for each row.
>    ct_expand:term(
>        [[[X,Y] || X <- [R], Y <- "123456789"] || R <- "ABCDEFGHI"]
>    ).
> box_squares() ->
>    %% All the square names for each box.
>    ct_expand:term(
>        [[[X,Y] || X <- R, Y <- C] || R <- ["ABC", "DEF", "GHI"], C <- ["123", "456", "789"]]
>    ).
>
>
> ________________________________________
> From: Ahmed Omar []
> Sent: Saturday, March 26, 2011 4:39 PM
> To: Evans, Matthew
> Cc: Andreas Pauley; 
> Subject: Re: [erlang-questions 20] Re: A sudoku solver in Erlang compared to Python
>
> BTW, beside looking into reducing eleminations (which requires deep dive), if u do some time profiling using eprof u will find that most of the time spent actually in the cross functions
> sudoku:'-cross/2-lc$^1/1-1-'/4                       96558660  37.46  35024918  [      0.36]
>
> On Sat, Mar 26, 2011 at 9:33 PM, Ahmed Omar <<mailto:>> wrote:
> actually instead of doing
>
>
>
>    NonUniquePeers = shallow_flatten([S || S <- units(Square)]),
>
>
>
>    PeerSet = sets:from_list(NonUniquePeers),
>
>
>
>    PeersWithSelf = sets:to_list(PeerSet),
>
>
>
>
>
>
>
>
> can't u just do
>
>
> PeersWithSelf = lists:usort(NonUniquePeers).
>
>
>
> ?
>
>
>
>
>
>
>
>
> On Sat, Mar 26, 2011 at 9:14 PM, Evans, Matthew <<mailto:>> wrote:
> Good call....gb_sets should be faster. Compiled to native it runs faster still
>
> 9> sudoku:print_results("top95.txt", "\n").
> Solved 95 of 95 puzzles from top95.txt in 0.810156 secs (117.26 Hz)
> (901201 total eliminations, avg 9486.33, median 6267, max 56820, min 1792).
> ok
>
> ________________________________________
> From: Ahmed Omar [<mailto:>]
> Sent: Saturday, March 26, 2011 3:08 PM
> To: Andreas Pauley
> Cc: Evans, Matthew; <mailto:>
> Subject: Re: [erlang-questions 20] Re: A sudoku solver in Erlang compared to Python
>
> using gb_sets instead of sets could decrease eliminations a bit and give u some boost. However, i didn't dive deeper into the code or the algorithm.
>
> On Sat, Mar 26, 2011 at 9:41 AM, Andreas Pauley <<mailto:><mailto:<mailto:>>> wrote:
> Thanks, I've changed the compile options and this definitely makes it
> somewhat faster:
> https://github.com/apauley/sudoku-in-erlang/commit/b7ce2abb6ca013850ed8f3e8fd7f5f6be7004cbb
>
> Strangely the native flag is not documented here:
> http://www.erlang.org/doc/man/compile.html#file-2
>
> The more interesting improvement would be to decrease the number of
> eliminations performed, but for that I'll have to go deep into the
> code :-)
>
>
> On Sat, Mar 26, 2011 at 12:16 AM, Evans, Matthew <<mailto:><mailto:<mailto:>>> wrote:
>> Without going deep into the code, one thing to try is compile with the native flag (running the tests from the VM shell here):
>>
>> Without native set:
>> 18> c(sudoku).
>> {ok,sudoku}
>> 19>  sudoku:print_results("top95.txt", "\n").
>> Solved 95 of 95 puzzles from top95.txt in 2.038825 secs (46.60 Hz)
>> (922678 total eliminations, avg 9712.40, median 6596, max 55370, min 1797).
>> ok
>>
>>
>> With native set:
>> 20> c(sudoku,[native]).
>> {ok,sudoku}
>> 21>  sudoku:print_results("top95.txt", "\n").
>> Solved 95 of 95 puzzles from top95.txt in 1.613416 secs (58.88 Hz)
>> (922678 total eliminations, avg 9712.40, median 6596, max 55370, min 1797).
>>
>>
>>
>> -----Original Message-----
>> From: <mailto:><mailto:<mailto:>> [mailto:<mailto:><mailto:<mailto:>>] On Behalf Of Andreas Pauley
>> Sent: Friday, March 25, 2011 8:15 AM
>> To: <mailto:><mailto:<mailto:>>
>> Subject: [erlang-questions 4] A sudoku solver in Erlang compared to Python
>>
>> Hi all,
>>
>> In my quest to learn Erlang I've translated Peter Norvig's sudoku
>> solver into Erlang:
>> https://github.com/apauley/sudoku-in-erlang
>>
>> For comparison, my slightly modified version of Norvig's Python code
>> can be found here:
>> https://github.com/apauley/sudoku-by-norvig
>>
>> I like the pattern matching, it was fun to do the entire solution
>> without a single if or case statement :-)
>> Something that bothers me though is that the Python solution seems to
>> be faster, even after I've made the Erlang solution use multiple
>> processors.
>>
>> In order to compare the two solutions I counted the number of
>> eliminations each performed.
>> The eliminate function is the core of the solution, for example
>> assigning a single value to a square is implemented as the elimination
>> of all other values.
>>
>> With the Erlang implementation I get:
>> sudoku-in-erlang$ ./sudoku
>> All tests passed :-)
>> Solved 50 of 50 puzzles from easy50.txt in 2.890978 secs (17.30 Hz)
>> (93403 total eliminations, avg 1868.06, median 1810, max 2517, min 1770).
>> Solved 95 of 95 puzzles from top95.txt in 22.004369 secs (4.32 Hz)
>> (922678 total eliminations, avg 9712.40, median 6596, max 55370, min 1797).
>> Solved 11 of 11 puzzles from hardest.txt in 0.851678 secs (12.92 Hz)
>> (32339 total eliminations, avg 2939.91, median 2894, max 4779, min 1781).
>>
>> And with the Python implementation:
>> sudoku-by-norvig$ ./sudoku.py
>> All tests pass.
>> Solved 50 of 50 puzzles from easy50.txt in 0.792008 secs (63.13 Hz)
>> (33059 total eliminations, avg 661.00, median 648, max 830, min 648).
>> Solved 95 of 95 puzzles from top95.txt in 5.903875 secs (16.09 Hz)
>> (221997 total eliminations, avg 2336.00, median 1492, max 11512, min 648).
>> Solved 11 of 11 puzzles from hardest.txt in 0.237532 secs (46.31 Hz)
>> (9436 total eliminations, avg 857.00, median 817, max 1198, min 648).
>>
>> So according to the stats above, the Python solution performs less
>> computations when given exactly the same input.
>> The Erlang code is as close to the Python as I could make it, I've
>> done more or less a direct translation of the algorithms used.
>>
>> I suspect that there are some lazy evaluation happening in the Python
>> version, possibly generators, although I haven't pinpointed it yet.
>>
>> How can I improve my Erlang code in this solution?
>>
>> Kind regards,
>> Andreas
>> _______________________________________________
>> erlang-questions mailing list
>> <mailto:><mailto:<mailto:>>
>> http://erlang.org/mailman/listinfo/erlang-questions
>>
> _______________________________________________
> erlang-questions mailing list
> <mailto:><mailto:<mailto:>>
> http://erlang.org/mailman/listinfo/erlang-questions
>
>
>
> --
> Best Regards,
> - Ahmed Omar
> http://nl.linkedin.com/in/adiaa
> Follow me on twitter
> @spawn_think<http://twitter.com/#!/spawn_think>
>
>
>
>
> --
> Best Regards,
> - Ahmed Omar
> http://nl.linkedin.com/in/adiaa
> Follow me on twitter
> @spawn_think<http://twitter.com/#!/spawn_think>
>
>
>
>
> --
> Best Regards,
> - Ahmed Omar
> http://nl.linkedin.com/in/adiaa
> Follow me on twitter
> @spawn_think<http://twitter.com/#!/spawn_think>
>
> _______________________________________________
> erlang-questions mailing list
> 
> http://erlang.org/mailman/listinfo/erlang-questions




More information about the erlang-questions mailing list