A Comparative Study of the Performance of Concurrency Control Algorithms in a Centralised Database

Sammanfattning: "The performance of Concurrency Control Algorithms (CCAs) in databases has been widely studied over the last ten years, mainly using analytic and simluation techniques. Results have been contradictory. The main CCA categories to emerge are locking, timestamping and serial validation. Most studies have considered one or two CCAs within the same framework and only one simulation study has considered all the categories.Experimental studies are rare and none has covered all three categories of CCA. In this study, the performance of atomic static locking (PRE), two-phase exclusive (2PLE) and update (2PLU) locking, basic timestamping (BTO) and serial validation (SV) is compared using a prototype database system. Previous studies suggest that there is little difference in CCA performance at low levels of conflict. This study explores a "worst-case" scenario and uses update only transactions while recognising that such an approach may be biased against CCAs which allow share access such as 2PLU. Benchmark results are presented for single user transactions. Two transaction types are selected for the multi-user experiments and are run under each CCA and without concurrency control to provide a baseline for comparison. An adaptive restart technique is introduced for 2PLU and BTO and is shown to considerably improve performance.The experiments are repeated using a simulation model and the adaptive restart technique is further investigated. Comparing results with the prototype and other simulation studies indicates that the effect of both blocking and restarting delays are important in determining relative CCA performance"