Data Oriented Transfer service is an architecture that aims to separate the data transfer from the control layer. This is related to PANTS, my class project. Even the APIs they have are similar to what we came up with.
I like the motivating scenarios in the paper - all of them seem practically useful to me. The authors come up with a solution that is flexible but obviously this seems like a presentation of a design space as opposed to any concrete solution. I like the way this paper goes about its job.
Their evaluations seem almost apologetic to me - it is more to show that performance does not really crap out. I would have liked a better, if only analytical, evaluation of the possible benefits of say, cross-application caching. They could have used some sample traces from a set of users to see if there are benefits to doing this. I am sure people will cricitize this paper on this count...and that would be sad.
I give them credit for integrating DOT with Postfix - for what it's worth, I always love a real implementation in a paper, even if it's novelty value isn't super-high. To be able to prove that things can work with your system is quite a big statement.
I vote for keeping it in.