130 likes | 259 Vues
This document summarizes the discussions from the DBMS Energy Awareness and Management Group, featuring contributions from experts like Michael Bender and Goetz Graefe. It highlights key considerations for optimizing energy efficiency in database management systems (DBMS), comparing it to traditional optimization for storage and performance. Topics include energy-efficient indexing, concurrency control, query execution strategies, and the integration of energy costs into user interactions and benchmarks. The aim is to explore avenues for cross-layer optimization and the impact of energy-aware practices in data management.
E N D
Inside the DBMS Energy Awareness and Energy Management
Group Participants • Michael Bender, Stony Brook University • Goetz Graefe, HP Labs • Le Gruenwald, National Science Foundation • Volker Hoefner, University of Kaiserslautern • SamirKhuller, University of Maryland • Bradley Kuszmaul, MIT • Alexandros Labrinidis, University of Pittsburgh • Mohamed Mokbel, University of Minnesota • MeikelPoess, Oracle Corporation • YichengTu, University of South Florida • Bo Zeng, University of South Florida
Repeated question • How is energy efficiency different than optimizing for • space (i.e., storage) and • time (i.e., performance)?
Indexing / Storage • How to build/maintain an index in an energy-efficient way? • E.g., deferred maintenance to handle spikes • Traditional trade-offs different now: • Load balancing VS switching off • How to consider different technologies at the same time? • Should we expose APIs for cross-layer optimization?
Concurrency Control & Recovery • Recovery/resiliency/fail-over are prime candidates for revisiting for energy efficiency • E.g., use (redo-only) logs vs copies • Power-aware concurrency control is possible • E.g., use latches more often / optimistic CC • Consider different storage layers/hardware alternatives
Query Execution (1) • What makes an algorithm energy-efficient? • Can new join/group-by algorithms be energy-efficient? • Is fast automatically energy-efficient? • No; E.g., Differential Voltage Scaling • Would data compression help? • More data fit in memory • Computation directly on compressed data?
Query Execution (2) • How can scheduling help? • E.g., load shaping by shifting load for later to avoid spikes (i.e., over-provisioning)
Query Optimization (1) • Cost models for energy consumption (need instrumentation) • Compile-time decisions should be shifted to run-time (to handle load/energy cost) • Binary decisions VS gradual transitions • Take into account different hardware options
Query Optimization (2) • Performance improvements: • percentage [not interesting] • factors [starts to get interesting] • orders of magnitude [really interesting] • ARE THERE ORDER OF MAGNITUDE OPPORTUNITIES? • Easy: utilizing new hardware • Difficult?
Query Optimization (3) • How to consider energy, as part of self-managing data management? • Auto-admin-style optimizers for storage/performance • Can there be query optimizers for energy consumption? • Can there be a “here’s 2 KWh, do the best you can” optimizations?
Benchmarks • Infrastructure needed to reduce barrier of entry to research in the area • E.g., a resource sharing repository as a start • Can we include energy in a way similar to $ for TPC benchmarks? • Idea for SIGMOD programming content topic to be energy-efficient algorithms
Involving the user (1) • Link query execution to energy spent • Use real dollar cost of energy instead of just amount of energy spent • Distinguish processing during peak energy demand hours VS low demand • Can differentiate for sustainability (i.e., charge for energy from renewable sources is cheaper) • Consider as part of SLAs (with big grain of salt)
Involving the user (2) • Vendors providing differentiated service, that includes energy costs • how can users optimize over different vendors?