| OCR Text |
Show 21 programs, e.g., compilers, is eligible for load migration. A load average, which represents the degree of idleness, is maintained in each network machine's kernel. Before a command is processed, the load averages from "every" network machine are obtained and the one with the minimum load average is chosen. The system then establishes the connections and sends the new task to this machine. Kratzer [40] suggested a swap bid protocol for distributed load balancing. When a processor receives a status update message, it finds, via a performance estimation heuristic, the best possible task movement to/from another node or a swap of tasks. Then it sends a swap bid that may or may not be accepted. The swap bid algorithm suffers two significant drawbacks: possible deadlock and status synchronization. Ni [46] proposed an attractive load balancing method for a small scale point-to-point multiprocessor system. An idle processor sends a request message to its neighbors. The neighbors respond with busy or not-busy status. The idle processor then selects a target neighbor and sends a draft message. The drafted processor may either respond with a new task or respond with a too-late message if the new task has been drafted by another processor. Tasks can be migrated at most one hop away from the originating host. The scheme is attractive for small-scale mult-iprocessor systems. The lack of provision for global load balancing limits the ability of the system to spread work beyond immediate neighbors of the originating node. . Stankovic [56] sketched a scheduling algorithm for a distributed computer . system. The algorithm is heuristic. which acts like an "expert system." He described general aspects of the scheduling problem and he hopes to integrate many of these requirements into an expert system program. This is a broad and ambitious approach. To a certain extent, many proposed algorithms, including one in this writing, can be considered to fall under this generalized scheme. No details of the "expert system" has been described up to now. The stochastic approach to solve the load balancing problem [11] assumes |
