Browse Prior Art Database

Early Commit Scheme for Multi-System Data Sharing Using Write Completion Log

IP.com Disclosure Number: IPCOM000040356D
Original Publication Date: 1987-Nov-01
Included in the Prior Art Database: 2005-Feb-02
Document File: 3 page(s) / 16K

Publishing Venue

IBM

Related People

Lee, YH: AUTHOR [+3]

Abstract

This invention is a method for early commit transactions in a multi-system data-sharing environment that allows for the early release of resources used by the transaction and also the optimizations of deferred write and notification-on-conflict. Database integrity is preserved across system failures by using a write completion log. Background The use of extended memory and disk controller caches in mainframe systems has the potential of substantially reducing the number of disk reads per transaction. In this situation, transaction response time will be dominated by the time for writing updated blocks to disk, if the disk writes need to be done before transaction commit.

This text was extracted from a PDF file.
This is the abbreviated version, containing approximately 43% of the total text.

Page 1 of 3

Early Commit Scheme for Multi-System Data Sharing Using Write Completion Log

This invention is a method for early commit transactions in a multi-system data-sharing environment that allows for the early release of resources used by the transaction and also the optimizations of deferred write and notification-on-conflict.

Database integrity is preserved across system failures by using a write completion log. Background The use of extended memory and disk controller caches in mainframe systems has the potential of substantially reducing the number of disk reads per transaction. In this situation, transaction response time will be dominated by the time for writing updated blocks to disk, if the disk writes need to be done before transaction commit. In IMS data sharing, a transaction does not commit until its updates are written to disk, the buffer invalidation message sent out, and response received from all systems. Since every committed transaction has its updates on disk, there is no need to redo the updates of committed transactions.

To serialize the access of data from concurrent transactions, two kinds of locks can be used in a multi-system environment. Global lock, e.g., block lock, will be held in the system level for serializing operations among systems. Within a system, a local lock of usually finer granularity, e.g., record lock, is granted to an individual transaction. When a system fails, the global lock held by that system must be held to prevent other systems from accessing the locked blocks. These blocks are potentially overwritten during the database recovery performed by the failed system. Many resources like PST and local locks are held until the completion of disk writes, a source of performance degradation. In this invention a scheme is proposed in which the transaction commit is done at the end of the execution of the transaction (of course, it is necessary to wait for the write-ahead log to be forced out). In the proposed scheme, transaction commit will not have to wait for the disk writes of the updated blocks to complete. Local locks and the address space held by the transaction are released when it is committed. The global locks of updated blocks continue to be held by the system. As will be seen below, when the update to the disk is complete, a log entry is forced out to the log. Then, the global lock of the updated block may be released. This scheme allows the exploitation of deferred writes, where the write of the updated block may be delayed so that if other transactions on the same system update the block again, some writes may be saved. Potentially, the write operation and the release of the block lock may be delayed until another system makes a claim for the block, for example, by the use of the notification on conflict - coherency control protocol 2. Notice that the logging of write completion is necessary to confine recovery operations in the failed system. If the completion of disk writes is not lo...