Using Agents to Exploit Heterogeneous Parallelism on High Performance Computers

摘要

While homogeneous parallelism has traditionally been exploited on scaleable high performance computers (HPCs) for applications such as signal processing, parallelism in requiring data dependent processing applications is often difficult to predict and exploit with traditional methods. This paper describes the adaptation of 'agent-based' systems that have been investigated in other domains such as the Internet. An agent is a autonomous process that adapts to its environment to accomplish a specific task. An agent first discovers the location of needed information, and then either sends the data to a central location or spawns processes to process the information in place. A process is spawned by encapsulating its executable code and state information (including how and where to send the results) into a package sent to the host compute node. The process is then scheduled and executed according to its priority.