An Energy-Aware SQL Database System
Abstract
As enterprise data volumes continue to grow, the energy consumption of large-scale database systems has become an important concern alongside performance and reliability. Existing energy-efficient database approaches, including PRE-BUD, Dynamic Power Management, and GreenDB, reduce energy use through power-state management. However, these systems often rely on simple frequency-based heuristics for hot-data placement, which rank tables by individual access frequency. This strategy becomes less effective in multi-tenant and multi-workflow environments, where queries often access coherent groups of related tables rather than isolated high-frequency tables. This dissertation presents GreenDB_DM, an energy-aware SQL database system that uses workload-aware data mining to improve table placement and reduce unnecessary cold-node wakeups. The first component is GARMT, a grouping-based association rule mining approach for SQL workload analysis. GARMT partitions SQL query logs into fixed-size groups and applies a modified FP-Growth algorithm, GFP-Growth, to discover frequent table co-access patterns. By aggregating consecutive SQL queries before mining, GARMT reduces the sparsity of individual statements and enables more meaningful workload-pattern discovery. Experimental results show that grouping reduces pattern-mining runtime by up to 40% compared with ungrouped baselines while preserving rule quality. The second component is a complementary study of dynamic data-driven modeling in recommender systems. Although it addresses a different application domain, this study supports the dissertation's broader methodological argument that system efficiency can be improved by modeling temporal and structural behavior rather than relying only on static metrics. Using a network-based Susceptible-Infected model with time-varying transmission rates, the study shows how recommendation strategies produce different propagation patterns over time. The central contribution is GreenDB_DM, which integrates GFP-Growth-based pattern discovery with a greedy set-cover placement algorithm to maximize full-query hit rate. By identifying frequently co-accessed table groups and placing complete groups on the hot node, GreenDB_DM increases the likelihood that entire queries can be served without activating cold nodes. Evaluation results show that GreenDB_DM improves full-query hit rate by 5–15 percentage points over the strongest placement baseline and reduces total cluster energy by an additional 2.4–4.7% beyond GreenDB. Together, these contributions show that energy-aware SQL database systems should not only manage hardware power states, but also exploit workload structure. This dissertation establishes workload-aware data mining as a practical foundation for building more adaptive and energy-efficient SQL database systems.