|
ABSTRACTCarnegie Mellon, School of Computer ScienceMatching Database Access Patterns to Storage Characteristics Jiri Schindler, Anastassia Ailamaki, Gregory R. Ganger Carnegie Mellon University Today’s storage interfaces hide device-specific details, simplifying system development and device interoperability. However, they prevent database systems from exploiting devices’ unique performance characteristics. Abstract and device-independent annotations to existing storage interfaces can cleanly expose key device characteristics that improve performance and simplify manual tuning. By automatically matching access patterns to device strengths, a database storage manager can achieve robust performance even with workloads competing for the same storage resource. For example, disk-optimized accesses result in simultaneous improvement of up to 3x for DSS workloads and 7% for a competing OLTP workload. As another example, accesses to relational tables can take advantage of MEMS-based storage parallelism to achieve order of magnitude improvements in selective scans. FULL PAPER: pdf |