Tutorial: Testing a conjecture in parallel (draft)ΒΆ

In this tutorial, we illustrate how to test a conjecture in parallel on a multicore machine using the @parallel decorator.

See also

sage.parallel

Todo

expand and move to sage.parallel.tutorial?

For illustration purpose, we take a stupid conjecture, namely that the number \(n=49\) has no divisor in the range \(2,...,9\). Let us start with a little function that checks the conjecture on a given \(n\) and \(i\):

sage: def check_conjecture(i, n):
....:      return not i.divides(n)

Shoot, the conjecture is false:

sage: n = 49
sage: all( check_conjecture(i, n) for i in srange(2,10) )
False

We can for example recover the counter example with exists():

sage: n = 49
sage: exists( ((i,n) for i in srange(2,10)),
....:         lambda (i,n): not check_conjecture(i,n))
(True, (7, 49))

We want to find a counter-example in parallel. To this end, we define a variant of check_conjecture that checks the conjecture in parallel on a bunch of inputs:

sage: @parallel
....: def check_par(i, n):
....:     return not i.divides(n)

Here is how we can use it to test the conjecture in parallel for three pairs (i, n):

sage: list( check_par( [ (2,11), (3, 9), (4,7) ] ) )
[(((2, 11), {}), True), (((3, 9), {}), False), (((4, 7), {}), True)]

Each output is of the form (input, result), where input describes the arguments and optional arguments passed down to the python function. We now run the check for \(n=49\) and the range 2,...,9:

sage: n = 49
sage: sorted(check_par( (i,n) for i in srange(2,10) ))
[(((2, 49), {}), True), (((3, 49), {}), True), (((4, 49), {}), True), (((5, 49), {}), True), (((6, 49), {}), True), (((7, 49), {}), False), (((8, 49), {}), True), (((9, 49), {}), True)]

If we just want to know whether \(n\) has no divisor in the range, we can do:

sage: all( result for (input, result) in check_par( (i,n) for i in srange(2,10) ))
Killing any remaining workers...
False

Note the information message just before the answer: the computation was stopped as soon as a counter-example was found.

Let us print all counter-examples:

sage: for (input, result) in check_par( (i,n) for i in srange(2,10) ):
....:     if not result: print input
((7, 49), {})

Here is a powerful idiom to recover a counter-example:

sage: for (input, result) in check_par( (i,n) for i in srange(2,10) ):
....:     assert result
Traceback (most recent call last):
...
AssertionError

There is no output if there is no counter-example. Otherwise, an error is thrown:

sage: for (input, result) in check_par( (i,n) for i in srange(2,10) ):
....:     assert result
Traceback (most recent call last):
...
AssertionError

Then one can use the Python debugger to recover the counter-example by post mortem introspection on the stack:

sage: import pdb
sage: pdb.pm()                            # not tested
> <ipython console>(2)<module>()
(Pdb) p input
((7, 49), {})

This may sound a bit of an overkill, but this is a very general technique, and it is handy when the input is a complicated object that we want to explore. In particular, one can use the following hack to insert the counter-example in the global name space:

sage: pdb.pm()                            # not tested
> <ipython console>(2)<module>()
(Pdb) p input[0][0]
7
(Pdb) import __main__
(Pdb) __main__.my_counter_example = input[0][0]

Now we can play with it:

sage: my_counter_example             # not tested
7
sage: my_counter_example.divides(49) # not tested
True