[jira] [Commented] (UIMA-4329) Object-based CAS implementation proposal/prototype

["Nick Hill (JIRA)" <de...@uima.apache.org>]

    [ https://issues.apache.org/jira/browse/UIMA-4329?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14486825#comment-14486825 ] 

Nick Hill commented on UIMA-4329:
---------------------------------

I tried it out with uimaFIT - in particular ran the unit tests in uimafit-core and uimafit-examples against it. There were a few errors:
- One compilation error where {{FeatureStructureImpl}} is referenced. I've added an empty 'dummy' class with the same name for now to prevent this error (the actual impl isn't needed)
- Some test failures related to xcas deserialization. This was a one small bug where a check needed to be moved, now fixed
- Some tests failed which relied on the {{FeatureStructure}} toString format that I hadn't impl'd. I copied the toString logic from the current impl and these tests now pass
- There were 3 remaining failed tests, all for the same reason which looks "legitimate". The tests assume the {{select()}} methods always returns FS's in the same order for different supertypes, but these use {{FSCollectionFactory.create()}} which contains logic to use either {{cas.getAnnotationIndex(type)}} or {{cas.getIndexRepository().getAllIndexedFS(type)}}. The former of these will be in a deterministic order but the latter may be ordered arbitrarily

I've updated the attached files with "v0.2" versions which contain these fixes plus some other minor cleanup/refactoring.

h5. Note on test times
All the tests seemed to run at a similar speed or faster, with the exception of {{JCasUtilTest.testSelectCoverRandom()}} which was slower. This appeared to be due to the different sorted index impl approach, which I think is faster for some use cases but slower for others. It should not be hard to modify it to be segmented by type similar to the bag index impl / existing sorted impl. I did also test a new {{AnnotationIndex.subiterator(start,end)}} method which 'directly' returns an iterator over all spanned annotations i.e. avoiding acrobatics currently required - this actually made the test in question faster than the existing impl, but I didn't include this in the attached update. It would require a change to consuming code which wouldn't be compatible with the current impl (although that might be a nice method to add to {{AnnotationIndex}} in any case!)

> Object-based CAS implementation proposal/prototype
> --------------------------------------------------
>
>                 Key: UIMA-4329
>                 URL: https://issues.apache.org/jira/browse/UIMA-4329
>             Project: UIMA
>          Issue Type: Brainstorming
>          Components: Core Java Framework
>            Reporter: Nick Hill
>            Priority: Minor
>         Attachments: uima-core_obj-0.2.jar, uimaj-core_obj-0.2.tar.gz
>
>
> I have been experimenting with a simplified CAS implementation where each feature structure is an object and the indices are based on standard Java SDK concurrent collection classes. This replaces the complex custom array-based heaps and index implementations.
> The primary motivation was to make the CAS threadsafe so that multiple annotators could process one concurrently, but I think there are a number of other benefits.
> Summary of advantages:
> - Drastic simplification of code - most proprietary data structure impls removed, many other classes removed, index/index repo impls are about 25% of the size of the heap versions (good for future enhancements/maintainability)
> - Thread safety - multiple logically independent annotators can work on the same CAS concurrently - reading, writing and iterating over feature structures. Opens up a lot of parallelism possibilities
> - No need for heap resizing or wasted space in fixed size CAS backing arrays, no large up-front memory cost for CASes - pooling them should no longer be necessary
> - Unlike the current heap impl, when a FS is removed from CAS indices it's space is actually freed (can be GC'd)
> - Unification of CAS and JCas - cover class instance (if it exists) "is" the feature structure
> - Significantly better performance (speed) for many use-cases, especially where there is heavy access of CAS data
> - Usage of standard Java data structure classes means it can benefit more "for free" from ongoing improvements in the java SDK and from hardware optimizations targeted at these classes
> I was hoping to see if there's interest from the community in taking this further, maybe even as a replacement for the current impl in a future version of uima-core. There has already been some discussion on the mailing list under the subject "Alternate CAS implementation".
> I'm attaching the current prototype, which should support most existing UIMA functionality with the exception of:
> - Binary serialization/deserialization
> - C/C++ framework (requires binary serialization)
> - "Delta" CAS related function including CAS markers
> - Index "auto protection" (recent 2.7 feature)
> Note I don't mean to imply these things can't be supported, just that they aren't yet.
> Where these things aren't used it should be possible to try out the attached uima-core.jar as a drop-in replacement with existing apps/frameworks. An important caveat though is that any existing JCas cover classes will need recompiling with the new jar (but not re-JCasGenning).
> I'll also attach the code. I started by basically ripping out the CAS heaps, so there's a lot of code which is just commented out (e.g. in CASImpl.java). Lots of cleanup/tidyup is still needed, and theres various places which still need fixing for threadsafety (e.g. synchronization around some existing create-on-first-access logic.. this is separate to the indices though). But those things shouldn't affect existing usage. A convention I followed was not to rename modified classes (e.g. CASImpl), but where an equivalent impl was created from scratch I did give it a new name starting with "CC" (e.g. FeatureStructureImpl is now CCFeatureStructure). The cc stood for "concurrent CAS". I have kept it in sync with the latest compatible changes in the uima-core stream, apart from those related to the non-impl'd functions mentioned above.
> Most of the "valid" unit tests work. Some are tied to the internals and no longer apply, many don't compile because they use binary serialization and/or delta CAS related classes which I removed for the time being. Some others I had to generalize a bit because for example they assumed a specific order in places where the order should be arbitrary, and maybe some other similar reasons.
> md5 checksums:
> {{4fd19b5f804fe8d505f697240c8e0366 *uima-core.jar}}
> {{51826aa44111b7f6e1fa307393eda8f4 *uimaj-core_obj.tar.gz}}



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)