Re: [Help-glpk] Performance decrease with more powerful computer

[Andrew Makhorin]

> As Jeffrey say:
> "Sometimes small changes in numerics can make a big difference in how 
> the branches of an MIP problem are explored"
> 
> The FPU Control Word directly impact how glpk will explore branches.
> It can be different by thread. The default value is not the same if you 
> are using visual studio, gcc or java ...
> GLPK don't set the FPU control word.
> http://msdn.microsoft.com/en-us/library/e9b52ceh(v=vs.80).aspx
> 

Glpk doesn't assume any specific floating-point settings. However, the
following settings are preferred:

i) all temporary results are evaluated to maximal possible precision
(e.g. FLT_EVAL_METHOD = 2 for IEEE 754 model that corresponds to 64 or
112 bits of mantissa);

ii) correct rounding is used ("round to nearest").

Please note that glpk conforms to C89, so it is unable to change the
floating-point environment--the user's program should do that, if
necessary.