Memory System Performance of a Database in a Shared-Memory Multiprocessor

Although cache-coherent shared-memory multiprocessors are sometimes used to run commercial workloads, little work has been done to characterize how well they support these work-loads. In particular, we do not have many insights on the demands of commercial workloads on the memory subsystem of such machines. These issues are important because of the large market for commercial applications. In this paper, we examine the memory performance of a multiprocessor emulation of Postgres95, a public-domain database. Using address traces of up to 8-processor runs of TPC-B, OO1, and TPC-D, we nd three major performance problems: contention for synchronization, stall due to I/O, and misses on shared data. Our data reveals a synchronization overhead equal to 29% of the time, which can be partially removed with ner-grained locks. I/O system calls can stall workloads for a period comparable to the workloads' computation time. However, increasing the size of the buuer cache to hold the database in memory reduces the stall by over 7 times. Finally, although the stall due to misses on shared data represents a modest 1-8% of the time, it will probably increase with more processors and, unexpectedly, is caused mostly by misses on database metadata.