A method that combines global load balancing with dynamic task scheduling on a multiprocessor machine is presented. This method does not require prior knowledge of the run times of tasks, and is based on a rough grammar representing distributed computation is presented. Production rules are dynamically constructed when a concurrent program is run. The set of the rough grammar production rules is updated and rolled in a pipeline fashion, together with codes of the processes. This pipeline fashion of rolling the jobs defines the global job balancing. The rough grammar uses any operators and metrics, not only concatenation, inside its production rules. Performance parameters for the combination of a global load balancing and decentralized dynamic task scheduling are derived and compared with those for a statically scheduled multiprocessor. Based on these parameters, the decentralized methodology is shown to attain a much higher performance level and a highly improved fault tolerance.>

The necessary and sufficient condition for deadlock in a distributed system and an algorithm for detection of a distributed deadlock based on the sufficient condition are formulated. The protocol formulated, checks all wait-for contiguous requests in one iteration. A cycle is detected when a query message reaches the initiator. A wait-for cycle is only the necessary condition for the distributed deadlock. A no-deadlock message is expected by the query initiator to infer a deadlock-free situation if at least one wait-for cycle is present. A no-deadlock message is issued by a dependent (query intercessor) that is not waiting-for. No no-deadlock message implies a deadlock, and processes listed in the received query messages are the processes involved in a distributed deadlock. Properties of the protocol are discussed. The authors show that a replication of a requested higher-priority (or older) process can prevent a distributed deadlock (in a continuous deadlock treatment). A replication is shown to recover (in a periodical deadlock handling) a sequence of processes from an indefinite wait-die scheme.>