This optimization yields the greatest benefit when the coordi nator has no other remote subordinates. If it has other subordi nates, they must all vote YES before the coordinator can send its YES vote to the last agent. The prepare message can be sent in parallel to multiple subordinates so that their phase-one processing can occur concurrently. Communication with a last agent cannot overlap any other commit processing. Thus, the last-agent optimization that reduces message flows to one agent conflicts with the optimization inherent in preparing multiple agents concurrently. However, if messages to one of the remote partners involve long network delays (i.e connection through satellite)
the last-agent optimization provides significant savings. It is, for example, preferable to prepare the closely located partners (fast first phase) and reduce the communication required with the faraway partner to one slow round-trip message exchange.