In response to Justin's question (below) about whether I noticed differences
in predictions across the computer and operating system matrix.
Looking first at ESP-r we carry out a regime of testing before
we release which involves running through ~300 models/scenarios
to check the impact of proposed code or model changes. These
test runs are done on different computers.
Factors such as compiler switches tend to introduce noise in the order
of a Watt/hr or a fraction of a Watt/hr for a zone or surface energy
balance. There are often rounding errors between different compiler
versions and different operating systems. Across the matrix I
will be using the GCC/G++/GFORTRAN compilers but often there
are differences in the compiler version. Again these tend to result
only in a bit of noise in the predictions rather than the kinds of
differences that might lead users to made different design decisions.
Exceptions are models that have some stochastic in nature
or require iteration - CFD runs, mass flow predictions
and the like occasionally diverge because, for example a control
switches a time step earlier or later. I would need to look more
closely at the outputs to provide a more definitive view.
In the case of EnergyPlus I was primarily looking at the run times
rather than checking for differences in the predictions. I just
randomly looked at a few files for the medium complexity model
between a legacy laptop and a current Dell desktop (3-4 chip
generations apart) and find the *.csv (~48MB) and *.mtd files
are identical the *.eso (~60MB) is identical (except for date stamp)
and *.bnd *.dxf *eio *.end *.err *.rvaudit (except for the time taken)
are essentially the same for the CFT solver and for the Conduction
Finite Difference solver.
Between the ARM and Intel there seem to be some differences but
I need to rerun the ARM to generate
*.csv and *.eso files. There might have been a point difference
in the EPP version. Also the Intel implementations sometime use
the downloaded version rather than a local compile while I must
always build the ARM executables from scratch.
Regards Jon Hand
