Performs fitting for mathematical models encoded in CellML (www.cellml.org), using the NeSI (www.nesi.org.nz) infrastructure.
- Shameer Sathar, Research Assistant, University of Auckland (various bug fixes, box-based random logarithmic implementation, refactoring), 2015.
- David Shin, Research Assistant, University of Auckland (second iteration implementation, GA bugfixes), 2015.
- Gene Soudlenkov, Computational Scientist, University of Auckland (initial implementation in NeSI, MPI code), 2013-2014.
- Mike Cooling, Systems Biomedicine Group, University of Auckland (concepts, virtual experiments, general design, minor bug fixes, test cases).
We are grateful to the Aotearoa Foundation for providing funds to develop this piece of scientific software.
| Cores | Runtime (mm:ss) |
|---|---|
| 1 | 70:00 |
| 8 | 09:51 |
| 16 | 05:26 |
| 32 | 02:41 |
| 64 | 01:32 |
| 128 | 00:59 |
- mpiicpc: an Intel MPI C++ compiler to compile the codebase. Load onto Pan Cluster's build node by:
module load intel/ics-2013
Note: You must be logged onto one of NeSI's "build-nodes" to load module and build projects
- CellML-API
- AdvXMLParser
The first thing you should do is to configure the include paths in makefile: XMLPARSER_PATH and CELLML_PATH.
Set XMLPARSER_PATH to point directly to the AdvXMLParser directory.
Example:
XMLPARSER_PATH=/projects/uoa00322/AdvXMLParser
Configure CELLML_PATH to point directly to the CellML-API directory.
Example:
CELLML_PATH=/projects/uoa00322/cellml-sdk
Compile the project with the makefile:
make
which should generate the binary cellml-fitter in your current directory.
Test the program with some supplied problems in the tests sub-directory:
Jump into any available test folder:
Example: IP3 model
cd tests/3-ip3model
Feel free to look into the example test (XML) and job (Slurm) files.
Before batching any job with the example Slurm files, make sure you configure the library paths. Since CellML compiler relies on GCC compiler to build the code, LIBRARY_PATH and LD_LIBRARY_PATH should be configured by pointing to CellML library directory. For example:
LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/projects/uoa00322/cellml-sdk/lib/
LIBRARY_PATH=$LIBRARY_PATH:/projects/uoa00322/cellml-sdk/lib/
Finally, batch the slurm job file. As an example, running sbatch short.sl will batch the example test short.xml on a single processor.
The verbosity of the output can be set by adding or removing "-v" from the "srun" command
srun ../../cellml-fitter test.xml -v -v
| Number of "-v"s | Output |
|---|---|
| 0 | Only the best overall population member is reported |
| 1 | As 0 plus each generation's best is reported |
| 2 | As 1 plus each generation's members are reported |
| 3 | As 2 plus results of each genetic operator are reported and fitness evaluation by target point |
| 4 | As 3 plus detail of each genetic operator is reported, and detail of each target point evaluation |
Quickest way is by looking at an example: adapted from short.xml in IP3model problem...
<?xml version="1.0"?>
<CellMLTimeSeriesFit>
<GA InitialPopulation="10" Generations="10" Mutation_proportion="0.4" Crossover_proportion="0.30" RNG="1">
<Alleles>
<Allele Name="kf5" LowerBound="1.0e-8" UpperBound="9.999e2"/>
<Allele Name="kf4" LowerBound="1.0e-8" UpperBound="9.999e2"/>
...
<Allele Name="Rpc" LowerBound="1.0e-2" UpperBound="5e3"/>
</Alleles>
</GA>
<VirtualExperiments>
<VirtualExperiment ModelFilePath="ip3model.cellml" Variable="IP3" ReportStep="50.0">
<!-- note this VE doesn't have any Parameters to set, but next VE does (ie optional Parameters entity here) -->
<AssessmentPoints>
<AssessmentPoint time="100.0" target="0.026761882" />
<AssessmentPoint time="200.0" target="0.032711469" />
...
<AssessmentPoint time="10000.0" target="0.015490316" />
</AssessmentPoints>
</VirtualExperiment>
<VirtualExperiment ModelFilePath="ip3model.cellml" Variable="Ca" ReportStep="50.0">
<Parameters>
<Parameter ToSet="Ls" Value="5.0"/>
</Parameters>
<AssessmentPoints>
<AssessmentPoint time="100.0" target="0.1" />
...
<AssessmentPoint time="10000.0" target="0.099907954" />
</AssessmentPoints>
</VirtualExperiment>
<VirtualExperiment>
...
</VirtualExperiment>
...
<VirtualExperiment>
...
</VirtualExperiment>
</VirtualExperiments>
</CellMLTimeSeriesFit>Each VirtualExperiment must have the ReportStep set. This is used by the simulator to force the integrator to report results at given time interval. If it is set to "0", it defaults to intervals of 1 time unit.
Load distribution is based on the workitems list. The list is populated with the item which contain data to be computed and an opaque key, which is the context of the requestor. Once the list is fully populated, call to process runs through the list requesting compute done on the data by sending the requests to the rest of the MPI ranks. Once the results are back, the observer function is called for each result passing the workitem and the result to the callback.
Fitness evaluation currently treats every target data point as of equal importance. This will result in biasing the fits to those experiments with more data targets, unless each experiment has the same number of data targets. If this becomes a problem in practice, we could implement a simple weighing system to resolve it.