Efficient locking and caching of data in the multisystem shared disks transaction environment

Abstract

This paper describes a technique for use when multiple instances of a data base management system (DBMS), each with its own cache (buffer pool), can directly read and modify any data stored on a set of shared disks. Global locking and coherency control protocols are necessary in this context for assuring transaction consistency and for maintaining coherency of the data cached in the multiple caches. The coordination amongst the systems is performed by a set of local lock managers (LLMs) and a global lock manager (GLM). This typically involves sending messages. We describe a technique, called LP locking, which saves locking calls when the granularity of locking by transactions is the same as the granularity of caching by the cache manager. The savings are gained by making the LLMs hide from the GLM the distinction between a transaction lock, called the L lock, and a cache-ownership lock, called the P lock, for the same object. The L and P locks for an object, though distinct at an LLM, are known as a single lock at the GLM. An LLM can grant an L or P lock request on an object locally if the combined lock mode of the L and P locks already held on that object by that LLM is equal to or higher than the requested mode. Such optimizations save messages between the LLMs and the GLM. Our ideas apply also to the client-server environment which has become very popular in the OODBMS area and to the distributed shared memory environment.