Oracle International Corporation

Ravi Prem Mirchandaney

Gary Hallmark

Roger J Bamford

Boris Klots

Jeffrey Fischer

A method and apparatus for distributing work granules of a parent task among processes running on various nodes in a multi-processing computer system is provided. The parent task is divided into work granules of varying sizes based on the location of the data that must be accessed to perform the work granules. At least one process on each of the nodes that will be assisting in the execution of the parent task is initially assigned a work granule based on efficiency considerations. Such efficiency considerations may include, for example, the location of the data to be accessed relative to the process, the current I/O load of the devices on which data to be accessed is stored, and the relative sizes of the various work granules. When a process completes the work granule assigned to it, the process is assigned one of the remaining unassigned work granules. Again the work granule assignment is made based on efficiency considerations. This process continues until all of the work granules have been completed.

Method and apparatus for distributing work granules among processes based on the location of data accessed in the work granules

A method and system are provided for reconfiguring a multiple node system after an epoch change in a manner that reduces the overhead and system unavailability typically incurred during reconfiguration. A resource-to-master mapping is established using the combination of a resource-to-bucket hash function and a bucket-to-node hash function. The resource-to-bucket hash function is not changed in response to an epoch change. The bucket-to-node hash function does change in response to epoch changes. Techniques are disclosed for adjusting the dynamic bucket-to-node hash function after an epoch change in a manner that load balances among the new number of nodes in the system. Further, the changes to the bucket-to-node assignments are performed in a way that reduces the number of resources that have to be remastered. In one embodiment, only those resources that lose their masters during an epoch change are assigned new masters during an initial reconfiguration. Load balancing is then gradually achieved by migrating resources after the system has been made available. The old masters of resources forward access requests to new masters of resources once they have transferred the master resource objects for the requested resources. In addition, techniques are disclosed for migrating resources from a node in anticipation of a planned shutdown of the node.

Growing community of inventors

Menlo Park, CA, United States of America

Ravi Prem Mirchandaney