|
Abstract : |
In the last fifteen years, lock managers for regular disk-based database systems have seen little change. This is not without reason, since traditional memory-resident lock managers have always been much faster than disk-based database storage managers and disk-based database systems had few alternative design options. However, the introduction of memory-resident database systems has created both new requirements and new opportunities for better lock managers. We present the design of a lock manager that exploits the special nature of the memory-resident storage component in the Starburst experimental database system. To achieve a performance advantage over traditional lock managers, our lock manager physically attaches concurrency control meta-data to the database data itself, thereby making the meta-data directly accessible, rather than indirectly accessible via a hash-table structure. Furthermore, our lock manager eliminates intention locks and changes the lock granularity of each relation dynamically to achieve high performance while ensuring a high level of sharing. Performance experiments comparing our lock manager with the default Starburst lock manager show that these basic design choices can provide significant performance gains. 1, |
