Fusion is useful for distributed agents that communicate within and across different machines. For closely coupled cooperating agents, communication overheads can constitute significant portions of their operating costs. For example, in the sample applications described above, when an `upstream' agent filters out much of the data, then the remaining data handed off to the 'downstream' agent may not result in significant communication-based overheads. This is the case for the `statistics' agent operating on the atmospheric data in the ISDA application, for example. Conversely, when such filtering does not remove much data, communication overheads may be substantial, as evident for many of the PSSPS application's agents.
The intent of agent fusion is to remove communication overheads from collaborating sets of agents and to enable optimizations across agent boundaries. Briefly, these overheads include actual network transport and protocol processing times, data copying costs, and thread/process scheduling and context switching costs that arise when tasks are performed by multiple vs. single agents residing on the same machine. Possible optimizations across agent boundaries are similar to those performed by compilers across procedure boundaries, including inter-procedural analysis leading to code or data motion and procedure integration where multiple overlapping procedures are combined into a smaller number of more efficient, combined procedures[3, 31, 32]. One particularly unfortunate communication cost is that incurred by multiple cooperating agents residing on the same machine, where their shared location is due to unforeseen agent migration actions (e.g., both agents `found' interesting data on the same machine, or both agents moved to that machine due to local resource availability). In this case, it is clear that such agents would operate much more efficiently if they were placed into the same address space and used shared user-level threads, as this would reduce agent invocation costs to the costs of a few procedure calls (via adaptors). It would also eliminate overheads associated with the implementation of asynchronous invocations, such as the use of additional threads and their scheduling and synchronization, and additional data copying due to asynchrony and kernel/user space crossings.
In summary, while agent morphing specializes individual mobile agents, fusion performs runtime optimizations across multiple, cooperating agents. Fusion may be applied repeatedly, to create single efficient agents from multiple collaborating agents.