Function.map() and parallel evaluation of gradient and Hessian

[Mikhail Katliar]

Dear All,

Let's say I have a NLP which looks like following:

x = MX.sym('x')

y = MX.sym('y')

f = Function('f', [y, x], [some_function(y, x)])

N = data.shape[1]

f_map = f.map('f_map', N, 'openmp', [1], [])

nlp = {'x' : x, 'f' : sum2(f_map(data, x))}

solver = nlpsol('NLPSolver', 'ipopt', nlp)

solution = solver()

The question is: will I have parallel evaluation of the gradient and Hessian inside the solver?

Best regards,

Mikhail

[Joris Gillis]

Hi Mikhail,

In principle, yes.
However, since reduced outputs are unimplemented for openmp map, it will switch back to serial map when reduced outputs are needed.
The reverse mode of a map with reduced inputs has reduced outputs...

The workaround is not use reduced inputs yet; just repmat'ing the input.

Best,
  Joris

[Mikhail Katliar]

I actually get this message during the solver construction:

CasADi warning: "OpenMP not yet supported for reduced outputs. Falling back to serial mode." issued  on line 99 of file "/home/kotlyar/software/casadi/casadi/core/function/map.cpp". 

How can I benefit from parallelization for objective functions like this?

суббота, 2 июля 2016 г., 19:44:53 UTC+2 пользователь Mikhail Katliar написал:

[Joris Gillis]

Right, that's the warning for the problem I mentioned.

Try something like (untested):

f_map = f.map('f_map', N, 'openmp')

nlp = {'x' : x, 'f' : sum2(f_map(data, repmat(x,1,N) ))}

Best,
  Joris

[Mikhail Katliar]

Hi Joris,

your solution seems to parallelize the execution (tested on small problem), but eats up all memory (on real-size problem). In my case, data has size 16384x636 and x is 30x4. When the solver is started, memory usage goes beyond 8Gb and the system runs out of memory. If I reformulate the problem without map(), memory usage is negligible, but there is no parallelization.

How can I make CasADi calculate my objective and derivatives in parallel without running out of memory in this case?

суббота, 2 июля 2016 г., 20:16:19 UTC+2 пользователь Joris Gillis написал:

[Joris Gillis]

Hi Mikhail,

This is gonna sound like deja-vue :-)

This is another known and open issue. Joel did some heavy lifting with new features for thread-safe memory management, but this hasn't been applied to openmp map yet.

For now, the openmp map implementation allocates  times the memory for a single function evaluation.

You may try a simple workaround: Divide up and in a large and small factor, and stack map serial and map openmp together:

f_map = f.map('f_map_inner', 2048, 'serial').map('f_map_outer',8,'openmp')

Best,

  Joris