Hi Thomas,
I'm sorry I don't have time right now for a proper response (buried under paper deadlines). There are certainly times when one will be faster then the other and the reasons are quite complicated. To complicate matters further it is very difficult to get benchmarks that don't lie about performance in this space. There are also alternative implementations that change the balance drastically. The only broad advice I can give is to benchmark the target application with both implementations to see how all the implications fall out. A broad description of the differences in implementation would be that MVars have a fairness guarantee (that does not come for free) for waking waiting threads. STM does not have this fairness which can lead to problems for programs that have quick transactions that always win over occasional long transactions (there are ways to avoid with a different implementation or with the cost of shifted to the programmer). My guess is in your particular benchmark the unfairness of STM works to your advantage and all the work is happening sequentially while the MVar version's fairness incurs frequent cache misses.
Ryan