【摘要】
This paper presents the design of the Concert2 loadrnsharing scheme for loosely-coupled distributed systems.rnA popular trend in load sharing is to use limitedrnamounts of information in scheduling. Existing loadrnsharing schemes are briefly discussed in the context of therntype and amount of information they use in scheduling.rnConcert2 is based on a detailed model of processrnbehavior and generates large amounts of information. Thernresource requirements of tasks are stored in the form ofrndetailed signatures that are updated each time a taskrncompletes. In this way the behavior of each type of task isrnlearnt. The load level at processing nodes is representedrnas a collection of load metrics. The load on the CPU,rnmemory, disk and network resources is recorded. Thernmechanisms that record this information have beenrncarefully designed to minimize the associated overheads.rnThe scheme is totally distributed. Through its richinformationrnbasis, Concert2 achieves automation andrntransparency to users and applications.rnThe development of Concert2 has been motivated inrnpart by the desire to carry out a detailed investigation ofrnthe costs and benefits of using different amounts ofrninformation, and using the information in different ways,rnin load sharing policies.