The logic behind the getChildren() calls is dictated by the JCR spec (e.g. SNS counting), so there isn't a way to turn that off (nor is it desired IMO).
However, in the case of external nodes we can look into optimizing the calls which, in the case of pages of children, load each page (like getChildCount(name)). One thing we can try is what you suggested - delegate these calls to the connector implementation. Please feel free to log an enhancement request describing this case.
Note: We cannot define external folder as empty as in this case after document is created or queried modeshape will fail with error stating that parent has no reference to it's child. Parent (folder) must still be able to provide it's children for reference validation.
Not sure I understand what you mean by this.
But for some cases like document creation getChildren is called anyways and walks through all the pages until gets to the end.
So it sounds like you've made the CmisConnector implement Pageable, which is great.
Also specified PageWriter.UNKNOWN_TOTAL_SIZE for children size to make modeshape use getChildReference when appropriate instead of getChildren operation.
That is correct. If you look at the logic here and here, returning PageWriter.UNKNOWN_TOTAL_SIZE for children size will make ModeShape call directly to the connector for the child reference by key and parent key, and this would be faster than going through all the segments.
But for some cases like document creation getChildren is called anyways and walks through all the pages until gets to the end. This is caused for example with LazyCachedNode.getChildReferences().getChildCount(name) operation when SNS is being checked before actual node creation. This action takes much time which increases with number of items.
So the question:
Is there any way to make modeshape not to call getChildren or minimize the frequency of these calls ?
Yes, there may be a way to improve this, as Horia mentioned above. Unfortunately, pushing more of this logic down to the connector means breaking changes to the connector SPI. That's something that we can do in the 4.0 timeframe (our current release effort), but is something we try to avoid in minor releases (e.g., 3.x). In 4.0 it also would have to compete with our other very high-priority items. Any assistance here would be greatly appreciated.
BTW, even though some of the methods seem like they are related to SNS, they aren't always used that way. For example, getting the number of children with a particular name does indeed allow us to make ModeShape behave properly with SNS, but it also allows us to find out whether there is a child with a given name (e.g., the method can return 0 or 1).
In the meantime, one thing that you can try is to play with the sizes of your pages. Increasing the size of the pages and thus returning fewer pages will mean your app consumes more memory, but it will require fewer calls to the connector. We can't really say what that is, since it depends so much on your particular environment (hardware, available memory, CPU utilization, network latency, etc.), your CMIS source, your application, and of course your data.
Also, can you add more detail in terms of the performance goals and actuals? How far off are you?