Highly Concurrent Locking in Shared Memory Database Systems

In parallel database systems, conflicts for accesses to objects are solved through object locking. In order to acquire and release locks, in the standard implementation of a lock manager small sections of the code may be executed only by a single thread. On massively parallel shared memory machines (SMM) the serialization of these critical sections leads to serious performance degradation. We eleminate the serialization by decomposing the complex database lock needed for granular locking into basic lock primitives. By doing so, we measured a speedup of a factor 200 on the SB-PRAM. Our method can be ported to any architecture supporting the used lock primitives, as most SMMs do.