bizarre SPMD error

I was using Matlab R2011b and had strange SPMD error.

The structure of the code looks like this:

load A.mat               % A is super large, e.g. 7GB

spmd

        B = ...; C = ...  % create some "small" local variable
        do something with A,B,C ...

end

Our machine has over 40GB memory.

Here are the problems:

1) The program runs perfectly well in parallel, if A is smaller, e.g 0.5G;

2) The serial version of the program runs perfectly well, if A is as huge as 7GB.

3) spmd reports error if A is huge:

---------------------------------------

Error using distcompserialize
Error during serialization

Error in spmdlang.RemoteSpmdExecutor/initiateComputation (line 82)
                fcns  = distcompMakeByteBufferHandle( ...

Error in spmdlang.spmd_feval_impl (line 14)
    blockExecutor.initiateComputation();

Error in spmd_feval (line 8)
        spmdlang.spmd_feval_impl( varargin{:} );

Error in myprogram >(spmd) (line xx)
        spmd

-----------------------

so this is a memory issue? But although A is huge, it is shared. My local variables are pretty small, a few MB.

Any idea what went wrong? Thanks.

Yingjie

Answer 1

[Walter Roberson]

Shared doesn't mean that copies of it do not have to be sent to the different workers (which, after all, might not be on the same host.)

Please see Edric's solution http://www.mathworks.com/matlabcentral/answers/20126-parfor-problems