Some of the problems addressed in the papar can be solved with compiler enforceable constraints such as is done in Haskell's type system.
What comes to the performance, the benchmarks in the paper range from no speedup at all with 8 cores to having 1.5x perf increase with 2 cores to 4x increase with 8 cores. To me, that definately sounds like something worth researching.
The point of that paper is that they have been researching it, quite extensively. But the performance results have often been disappointing, and the semantics of nested transactions can be surprisingly complex. At what point do you start looking into other things?
I think transactional memory systems with at least some hardware support are potentially more interesting.