I was using Cbc.

SolveModel is a copy and paste job from JuMP (from the last release rather
than master) so may not work with JuMP from master - I couldn't get the
version from master working since it was incompatible with the JuMP release
I had! It'd be great to just be able to just include the file, but I
couldn't get that working so I just pasted it in (I should probably clean
Bench as I made quite a mess, apologies about that.)... so it may be you
need to update SolveModel from JuMP master/your version of JuMP to get
Bench working.

It's amazing how some small tweaks like this go so far, there's a few other
bits that are obvious even to me (but I just couldn't get working).


On 1 July 2014 15:47, Iain Dunning <[email protected]> wrote:

> JuMP won't be getting any faster, its entirely limited by the speed of the
> MIP solver. Which one did you use?
>
>
> On Tuesday, July 1, 2014 6:47:04 PM UTC-4, Iain Dunning wrote:
>>
>> I was unable to run Bench.jl (ERROR: varzm! not defined), but, on my
>> computer just using runtests.jl, a fixed seed, and total time for 100 random
>>
>> *Initial
>> elapsed time: 1.641434988 seconds (282491732 bytes allocated, 5.99% gc
>> time)
>>
>> *Change globals to const
>> elapsed time: 1.563094028 seconds (261818132 bytes allocated, 6.61% gc
>> time)
>>
>> * Changing from using a Dict{Int64, *} for the Grid types to just a
>> Vector{*}, as well as those other globals
>> elapsed time: 1.373703078 seconds (191864592 bytes allocated, 4.91% gc
>> time)
>>
>>
>>
>>
>>
>>
>>
>>
>> On Tuesday, July 1, 2014 6:27:15 PM UTC-4, andy hayden wrote:
>>>
>>> Bench.jl has a bench_compare method which returns a DataFrame of times
>>> (I then divide the median of Python vs Julia columns), I'll add this output
>>> to the Bench script as it's useful to see (would be nice to add more stats,
>>> as it's just a DataFrame of all the solved puzzles in seconds). By default
>>> it runs a hundred random sudoku's on Julia, Python, and JuMP (the same on
>>> each)...
>>>
>>> Thanks Steven: Making those const makes a huge difference, Julia wins
>>> (from 20% slower to 10% faster for me with just that change).
>>> I will have a play and see how your other suggestions play out.
>>>
>>> I was also very impressed with JuMP here... and it may be the latest is
>>> even faster (I'm using the version from the last release rather than
>>> master, and it has changed since then).
>>>
>>>
>>> On Tuesday, 1 July 2014 15:11:27 UTC-7, Iain Dunning wrote:
>>>>
>>>> I'm working on improving this, but I'm not sure how you are measuring
>>>> that 20% slower - can you be more specific?
>>>>
>>>> On Tuesday, July 1, 2014 1:37:00 PM UTC-4, andy hayden wrote:
>>>>>
>>>>> I recently ported Norvig's Solve Every Sudoku Puzzle
>>>>> <http://norvig.com/sudoku.html> to Julia: https://github.com/
>>>>> hayd/Sudoku.jl
>>>>>
>>>>> Some simple benchmarks suggest my Julia implementation solves around
>>>>> 20% slower* than the Python version, and 3 times faster than the
>>>>> implementation on JuMP (vendorized from the latest release), against the
>>>>> random puzzles. I tried to include the solver from attractivechaos/plb
>>>>> <https://github.com/attractivechaos/plb/tree/master/sudoku> but
>>>>> couldn't get it working for comparison...
>>>>>
>>>>> I'm new to Julia so would love to hear people's thoughts / any
>>>>> performance tips!
>>>>> I've not delved too deeply into the Profile, but @time suggests 10% of
>>>>> time is GC.
>>>>>
>>>>> **I'm sure I've lost some performance in translation which could be
>>>>> easily sped up...*
>>>>>
>>>>> Best,
>>>>> Andy
>>>>>
>>>>

Reply via email to