Many kudos to the author for:
a) posting the row size
b) posting the environment used
c) running equivalent tests
d) running the tests several times in a row
e) choosing a large enough data set to override any potential caching that may have happened.
Very thorough test. Interesting results, and I can see how these would vary. What I would be interested in next, is to see these tests run on Windows64 with SQLServer 64 bit, where SSIS 64 bit would be tested against JVM 64 bit. I'm assuming these tests were run in 32 bit. Java seems to react better in a 64 bit environment.
On another note: while SSIS is interesting, it does have it's pains (as the author pointed out) with metadata and changes. I would recommend possibly checking out a tool called mapping manager from http://AnalytixDS.com which generates SSIS code (they are working on generating Talend Code) from a single set of metadata (cross-references/cross-walks). Anyhow, just a thought.
Thanks for the insightful read.
Hope this helps,