The performance analysis of concurrency control protocols based on speculation for distributed database systems

摘要

In this paper we investigate the performance of locking protocols based on speculation for distributed database systems (DDBSs). In speculative locking (SL), a transaction releases the lock on the data object whenever it produces the corresponding after-image during execution. By accessing both before- and after-images, the waiting transaction carries out speculative executions. As compared to locking, SL increases concurrency by allowing more parallelism among conflicting transactions without violating the serializability criteria. However, SL demands extra processing and main memory resources to support multiple executions for a transaction. By exploiting the fact that a transaction is more likely to commit than abort, we vary robustness of a transaction against cascading aborts and propose three variants of SL: SL(n), SL(1), and SL(2). Through simulation study we have compared the performance of SL variants with conventional two-phase locking, wait depth limited and optimistic locking approaches. The results indicate that even with manageable extra main memory and processing resources, these variants offer substantial benefits in DDBS (wide area network) environments where a transaction spends longer time in processing (execution and commit) and transaction-aborts occur frequently. Among SL variants, SL(2) performs closely with SL(n). SL(l) increases performance significantly over SL(2) under normal abort rate. However, under high abort rate, SL(2) improves performance over SL(1) as SL(2) is more robust against transaction aborts.