International Business Machines Corporation

Anupam K Bhide

Daniel Manuel Dias

Ambuj Goyal

Joel Leonard Wolf

Francis Nicholas Parr

Chandrasekaran Mohan

Hui-I Hsiao

George Prentice Copeland

Anant Deep Jhingran

Christos Alkiviadis Polyzois

A method and system are provided for partitioning a database and for balancing the processing load among processing nodes in a data processing system. A database partition is stored at each node and a logical owner node is assigned to each partition. The logical owner becomes responsible for all database requests related to the corresponding partition. Techniques for optimally assigning logical owners to all partitions in a system are provided and are responsive to parameters such as the load at each node, the combined load of the system, and the total number of assignments allowed. Techniques are also provided to accomplish the transfer of partition ownership between processing nodes when multiple transactions related to the partition are simultaneously occurring.

Method and system for database load balancing

A high performance disk subsystem has a disk write mechanism that provides high throughput for random disk writes by dividing the disks into a data portion and a log portion. A fault-tolerant memory in the disk controller is used to temporarily store data blocks or pages which are to be written in the data portion of the disks. The updated blocks are sorted in the fault-tolerant memory in accordance with their physical home location on the data portion of the disks. Sorted runs are temporarily written out to the log portion of the disk. In a parallel process, sorted runs are retrieved from the log portion of the disks, merged in the fault-tolerant memory, and written out to their home location on the data portion of the disks, which converts random disk writes to largely sequential I/O while retaining physical clustering of the data on disk.

Disk storage apparatus and method for converting random writes to

In a partitioned database system of the Shared Nothing type, one or more secondary replicas of each partition are maintained by spooling (i.e., asynchronously sending) modified (usually called dirty) pages from the primary replica to the secondary replica(s) rather than by using a synchronous page update or by sending log entries instead of entire pages. A Write-Ahead Log protocol is used so that a dirty page is not forced to non-volatile storage until a log record of the modification is created and written to non-volatile storage. Replica updating does not delay the committing of transactions because replica updating is done asynchronously with respect to transaction processing. Since dirty pages are sent rather than only log entries, disk accesses and processing at the secondary replica(s) arising from the maintaining of the replicas are minimized as well. Only one centrally accessible log is maintained for all replicas of the same partition.

Asynchronous replica management in shared nothing architectures

A fault-tolerant high performance mirrored disk subsystem is described which has an improved disk writing scheme that provides high throughput for random disk writes and at the same time guarantees high performance for disk reads. The subsystem also has an improved recovery mechanism which provides fast recovery in the event that one of the mirrored disks fails and during such recovery provides the same performance as during non-recovery periods. Data blocks or pages which are to be written to disk are temporarily accumulated and sorted (or scheduled) into an order (or schedule) which can be written to disk efficiently, which in a preferred embodiment is in accordance with the physical location on disk at which each such block will be written. This also generally corresponds to an order which is encountered by a writ head during a physical scan of a disk. The disks of a mirrored pair are operated out of phase with each other, so that one will be in read mode while the other is in write mode. Updated blocks are written out to the disk that is in write mode in sorted order, while guaranteed read performance is provided by the other disk that is in read mode. When a batch of updates has been applied to one disk of a mirrored pair, the mirrored pair switch their modes, and the other disk is updated. Preferably the updates are kept in a non-volatile memory, which furthermore advantageously may be made fault-tolerant as well. During recovery a pair of spare alternating mirrored disks is introduced to which new updates are directed, while a background scan process copies data from the surviving disk to the new mirrored pair.

Growing community of inventors

Foster City, CA, United States of America

Anupam K Bhide