This evening, Nguyen of my group found a better configuration of the Apache MQ system that does not try to store all the messages on disk (the default behavior for AMQ is to do that, in order to achieve guaranteed delivery). But we don't need guaranteed delivery for the server, only for the brokers, because if the server dies we have no way to resume a simulation anyway.
The bottom line is that running on my desktop machine, with the server running inside STS and the sample broker running under maven, the persistent version took on average 4300 msec/timeslot, and the non-persistent version took on average 218 msec/timeslot. This means that we should very rarely see pauses, and it gives us some additional slack to improve our models without starving the brokers for decision time.
The updated version is committed on the master branch.