Dissecting RIF/ReqIF metamodel
RIF/ReqIF is the new OMG standard for requirements interchange. RMF (currently in proposal phase) provides an EMF Ecore based metamodel implementation for the RIF/ReqIF XML format. The metamodel is a clean implementation of the format without any “XML noise”. The Ecore metamodel also conforms to the CMOF metamodel delivered by OMG (as it has been derived from it). The metamodel reads/writes RIF/ReqIF data in conformance with the RIF/ReqIF XML Schema.
The challenge in the Ecore based RIF/ReqIF implementation was the customization of loaders and serializers to make them RIF/ReqIF XML Schema conformant. EMF provides different ways of customizing the XML output.
1. Using ExtendedMetadata annotations.
Both approaches are quite tricky to implement when the expected XML output has structural differences compared to the Ecore metamodel. For example, the XML output has wrapper elements for lists (ELists) in the Ecore metamodel.
We went in for a third approach.
We imported RIF/ReqIF XML schema using EMF importer to create a RIF/ReqIF XML Ecore model (Anyone who has done this, knows it’s an “ugly” metamodel that EMF generates here). The next step was to create the real RIF/ReqIF metamodel, by importing RIF/ReqIF CMOF file. We then wrote a generic (of course reusable) Ecore XML to Ecore converter to do a model to model transformation in both directions. The whole processing is cleanly hidden in a new EMF Resource implementation such that the user hardly knows anything about these.
The following are the advantages we saw in the approach.
1. You don’t do customizations at a XML level, but at a higher EMF model API level. That is, you don’t hack some XML SAX events, but work with familiar EMF APIs.
2. The generic Ecore converter has enough hooks to plugin in the transformations.
3. Easy maintenance as the customizations are available in one place.
The drawback of this approach could be the processing involved in the model to model transformation. However, the highly optimized model to model transformation we implemented doesn’t make it all that slow. Our performance tests proves this (which would be the topic of my next blog post). It is infact upto 300 times faster than having the same data in the default XMI format (Yes I meant 300! Ed Merks is going to get back to me on this I hope ;) )
Ed did get back to me. With the performance optimizations applied to XMIResource loader to use cache for Intrinsic Ids and URIs, the performance improved dramatically. The XMI resource loader now takes only 22,2s to load a 32MB sample, compared to the earlier 70mins. I am glad that our RIF XMLResource however takes only 14,4s for the same contents.