Hector: an agent based architecture for dynamic resource management

– Networked workstations have emerged as an increasingly popular style of computation. Spurred by advances in computational capability and network bandwidth, they offer cost–effective computation. Because of the fine–grained, loosely coupled nature of the hardware, an agent–based system to monitor program performance can be used to harness the aggregate capability of workstations efficiently, rapidly, automatically, and with minimal intrusiveness. In the Hector distributed run–time environment, the agents that run on each candidate platform are called “slave allocators”, and they forward commands from the central allocation system, monitor system availability and loading, and can transparently extract detailed performance information from programs as they run. The overall architecture of the agent–based allocation system is detailed, as are testing and results that demonstrate its efficiency, low overhead, and utility. Measurements of the overhead of allocation, control, migration, and performance instrumentation are included.