On the Locality and Scalability of Database Workloads

Continuous advances in technology allow increased capacity and density for memories and disks. Consequently, the size of the applications currently limited by the availability of these resource will inevitable grow. This paper considers one such class of applications, namely database workloads, and investigates their locality and Scalability in two main dimensions; database working set hierarchy and the communication behavior when varying the database size and the number of processors accessing the database. We use the IBM DB2 database management system and the TPC-C and TPC-D benchmarks as our vehicle for experimentation.

We find that the first important working set sizes of the two benchmarks are independent of the database size. The database size, however, impacts directly on the second important working set sized. With the three selected TPC-D queries (Q3, Q6, and Q7), the second working set sizes increase (almost) as fast as the database grows. For TPC-C, a larger database receives a smaller incremental benefit from increasing cache size. When multiprocessors are employed, increasing the number of processors naturally increases the coherence traffic. When the number of processors is fixed an the per processor cache size is increased, the amount of coherence misses goes up slightly, but its relative significance rises quickly as the amount of capacity misses drops at the same time.