Faster model distribution

When working with a large model, it may happen that the time required to send the model to the worker nodes takes a significant portion of the total computation time.

You can use Julia serialization to prevent this. Because the shared filesystem is a common feature of most HPC installations around, you can very easily utilize it to broadcast a serialized version of the model to all worker nodes. In COBREXA, that functionality is wrapped in a Serialized model type, which provides a tiny abstraction around this functionality:

• you call serialize_model before you start your analysis to place the model to the shared storage (and, possibly, free the RAM required to hold the model)
• the freshly created Serialized model type is tiny – it only stores the file name where the model data can be found
• all analysis functions automatically call precache! on the model to get the actual model data loaded before the model is used, which transparently causes the loading from the shared media, and thus the fast distribution

The use is pretty straightforward. After you have your model loaded, you simply convert it to the small serialized form:

model = load_model(...)
# ... prepare the model ...

cachefile = tempname(".")
sm = serialize_model(model, cachefile)

Then, the analysis functions is run as it would be with a "normal" model:

screen(sm, ...)

After the analysis is done, it is useful to remove the temporary file:

rm(cachefile)