Please note: This PhD seminar will take place in DC 3301.
Zeynep Korkmaz, PhD candidate
David R. Cheriton School of Computer Science
Supervisors: Professors M. Tamer Özsu, Khuzaima Daudjee
Graph database management systems (GDBMSs) are increasingly used to support traversal-heavy workloads. Graph queries often follow data-dependent paths through the graph, producing accesses that do not align with the storage hierarchy, from data layout on disk to main-memory data pages. However, the structural characteristics of graphs can provide useful information about which vertices and edges are likely to be accessed together during query execution. When storage and caching mechanisms are agnostic to graph structure, they can fail to exploit opportunities for locality.
In this seminar, first, I briefly talk about how incorporating graph topology into storage and cache-layer design enables the leveraging of locality benefits. Then, I address a problem observed while designing the storage and cache-layer optimizations: the commonly used locality quality metrics do not always correlate with actual query performance. Finally, I present a systematic study of how these metrics fail to consistently predict query efficiency and explore the underlying reasons for this irregularity.