why does (ga + parallel + global variables + sub2ind) fail?

When optimizing with ga and parallel computing, a sub2ind call using globals fails (subscript vectors must be of same size), but when doing it in serial, it doesn't fail. Why?

Answer 1

[Matt J]

Probably because of the dangers of using global variables.

[Jochen Schuettler]

Hi Matt, of course I know all about this. There is a bunch of constant data for the fitness function to know. It would really slow down optimization if I unnecessarily copy all that data again and again as a function input. Could you please give a more helpful answer?

[Stephen23]

"It would really slow down optimization if I unnecessarily copy all that data again and again as a function input"

Not so much... MATLAB only copies variables in the memory when they are changed: "MATLAB internally optimizes for some cases and passes by reference and only creates local copies if modifications are being made to these arguments." source:

https://www.mathworks.com/matlabcentral/answers/96960-does-matlab-pass-parameters-using-call-by-value-or-call-by-reference

So as long as those variables are not being changed you can use them as inputs to a million function calls, and yet there will be no copies in MATLAB's memory.

[Matt J]

See also this page on preferred methods of passing around constant data.

[Walter Roberson]

Note that the above discussion is only for serial computing. The considerations are a bit different for parallel computing, where each worker must have a copy of the data because they are different processes.

[Matt J]

The discussion still applies to parallel computing if we're not talking about the one-time cost of broadcasting constant data to the workers. The OP seems to think that arguments passed to the fitness function will be copied "again and again" each time the fitness function is called. That is not true, even on parallel workers

Answer 2

[Walter Roberson]

Global variables are never copied to parallel workers.

You might be able to take advantage of parallel.pool.Constant or of parfevalOnAll() to initialize the variable on all of the workers.

[Walter Roberson]

If you have especially large shared data, you could also use the File Exchange contribution https://www.mathworks.com/matlabcentral/fileexchange/28572-sharedmatrix to use shared memory. This will only work if the compute nodes are on the same host, though

Answer 3

[Jochen Schuettler]

Thanks to everyone!