Speeding-up multiprocessors running DBMS workloads through coherence protocols

Abstract

In this work, it is shown how a DBMS workload, running on a shared-bus shared-memory multiprocessor, can be accelerated by adding simple support to the MESI coherence protocol. As a DBMS workload, we choose the TPC-D benchmark running on the PostgreSQL DBMS. Results show that, for a DSS workload, the use of a WU protocol with a selective invalidation strategy for private data improves performance because of the access pattern to shared data and the lower bus utilisation due to the absence of invalidation miss, when the contribution of passive sharing is eliminated. In the 16 processor case, the advantage can be quantified in a 20% of increased performance. Finally, it is shown how results can be extended to other DBMS workloads.