Source-linked AI summary
Energy Efficiency: The New Holy Grail of Data Management Systems Research
Stavros Harizopoulos, Mehul Shah, Justin Meza, Parthasarathy Ranganathan
TL;DR
Rising data-center energy costs and hardware limitations expose a gap in data-management software research. The paper examines software choices through two experiments and proposes general strategies and database-system opportunities for improving energy efficiency. It concludes that software will play a significant role alongside hardware and urges the community to prioritize energy-aware computing.
Problem
Rising energy costs and limited hardware power controls have received little attention from the data-management software perspective, despite growing data-center workloads.
Method
The paper discusses software-level opportunities, illustrates them with two database experiments, and organizes energy-waste reduction into three increasingly complex approaches.
Results
The examples show that database choices can improve energy efficiency independently of performance improvements, and that energy-efficient algorithms need not deliver the best performance.
Takeaways & Limitations
Data-management software will play a significant role in optimizing energy efficiency, with promising opportunities across physical design, resource management, query processing, and system architecture.
Abstract
from arXiv · showhide
Energy costs are quickly rising in large-scale data centers and are soon projected to overtake the cost of hardware. As a result, data center operators have recently started turning into using more energy-friendly hardware. Despite the growing body of research in power management techniques, there has been little work to date on energy efficiency from a data management software perspective. In this paper, we argue that hardware-only approaches are only part of the solution, and that data management software will be key in optimizing for energy efficiency. We discuss the problems arising from growing energy use in data centers and the trends that point to an increasing set of opportunities for software-level optimizations. Using two simple experiments, we illustrate the potential of such optimizations, and, motivated by these examples, we discuss general approaches for reducing energy waste. Lastly, we point out existing places within database systems that are promising for energy-efficiency optimizations and urge the data management systems community to shift focus from performance-oriented research to energy-efficient computing.
1. INTRODUCTION
Rising data-center energy use makes efficiency an important systems concern, but hardware improvements alone are insufficient. The paper uses database software choices and three classes of optimization to reduce energy waste.
- Motivation: Data-center power consumption is rising, increasing concerns about cost, density, scalability, reliability, and environmental impact.The paper cites a doubling of US data-center power consumption from 2000 to 2006 and another projected doubling within five years.
- Software opportunity: Hardware energy proportionality is only part of the solution; database software offers broad optimization opportunities through physical data independence and query optimization.Increasing heterogeneity across clusters, storage hierarchies, and CPUs further expands these opportunities.
- Evidence: Two examples show that physical design can improve energy efficiency and that performance optimization can differ from energy-efficiency optimization.The examples use an audited TPC-H-like configuration and a relational scan operator.
- Approaches: The paper proposes adjusting configuration and query-optimization parameters, consolidating resource use, and redesigning data structures, algorithms, and policies.These approaches are ordered by increasing complexity.
- Research agenda: Promising database research areas include physical design, storage and buffer management, query optimization and processing, new architectures, and special-purpose engines.The paper aims to raise awareness of energy-efficient data-management software as a research opportunity.
2. THE NEED FOR ENERGY EFFICIENCY
Growing data-management demand coincides with rising energy costs and limits imposed by power, cooling, and hardware controls. The paper argues that database systems can use workload-aware software choices across heterogeneous resources to improve energy efficiency.
- Foundations: Energy efficiency measures computing work per unit energy, and for fixed work it can improve by reducing power, time, or both.The paper relates energy to average power multiplied by execution time and describes efficiency as analogous to miles per gallon.
- Energy costs: Data centers used 60 billion kWh, or 1.5% of total US energy use, in 2006, and this use was expected to nearly double by 2010.The paper also reports $2.7 billion spent in the US and $7.2 billion worldwide to power and cool servers in 2005.
- System implications: Energy use constrains data-center density, reliability, scalability, and environmental impact because power and cooling capacity limit rack deployment and undercooled equipment has higher failure rates.Governmental agencies and industry groups were actively addressing enterprise power concerns.
- Workload pressure: Growing workloads span transaction processing, business intelligence, search, multimedia, web analytics, and cloud computing, motivating a shift from pure performance optimization toward energy efficiency.The paper notes continuing demand for cheaper, faster, and larger data-management systems.
- Hardware limits: Existing CPUs, memory, and disks provide limited power-performance control, and memory and disks may be either on at full power or off with expensive transitions.This limited dynamic power range restricts software’s ability to vary power consumption.
- Software opportunity: Database software can exploit physical data independence and query optimization to select hardware, consolidate data and processes, and choose algorithms that exercise components differently.Hardware heterogeneity is increasing across clusters and platforms, creating more power-performance choices.
3. EXAMPLE OPPORTUNITIES
Two experiments show that database choices can improve energy efficiency independently of performance: fewer disks can reduce energy use, while compression can speed queries yet increase energy consumption.
- Database choices can improve energy efficiency independently of improving performance.
- 3.1 Example 1: Diminishing Returns: The disk subsystem consumed more than 50% of total system power, making disk repartitioning an effective coarse power control.
- 3.1 Example 1: Diminishing Returns: 66 disks marked diminishing returns: energy efficiency rose 14% while performance fell 45%.The percentage performance gain no longer outweighed the percentage power gain.
- 3.2 Example 2: Algorithm Design: Compression reduced query time to 5.5 seconds from 10 seconds, producing a 2x speedup for the tested configuration.The compressed query spent 5.1 seconds in CPU time.
- 3.2 Example 2: Algorithm Design: 338 Joules for uncompressed data versus 487 Joules for compressed data shows that the slower query was more energy-efficient.The example assumes an idle CPU consumes no power, or that concurrent work uses the remaining CPU cycles.
- 3.2 Example 2: Algorithm Design: Energy-efficient algorithm design must evaluate power and performance together because their preferred choices can diverge.
4. APPROACHES FOR REDUCING WASTE
The paper proposes three increasingly complex software approaches for reducing wasted energy: tune existing controls, consolidate resource use, and redesign components or policies.
- The three software approaches increase in complexity from configuration tuning to resource consolidation to component redesign.
- Configuration Tuning: Tune system-wide knobs and query-optimization parameters for energy-efficient configurations on the underlying hardware.Relevant controls include parallelization, memory assignment, temporary space, and physical-design inputs.
- Configuration Tuning: Energy optimization requires revisiting tradeoffs such as CPU cycles versus disk-to-memory and memory-to-CPU bandwidth.Compression illustrates this tradeoff, and energy-oriented query plans may favor nested-loop joins over memory-intensive hash joins.
- Resource Use Consolidation: Consolidate computation and data across time and space so unused components can enter suspended or reduced-power modes.Software movement is worthwhile when its energy savings exceed the movement overhead.
- Resource Use Consolidation: Current components have limited power states, and switching costs can exceed the resulting energy savings.
- Resource Use Consolidation: Workload batching, bursty access, prefetching, caching, and data movement can lengthen idle periods or consolidate space-shared resources.These techniques may trade increased latency or data movement for opportunities to power down hardware.
- Software Redesign: Redesign software to minimize energy use, reduce code bloat, and accept selected sacrifices in other properties or metrics.Reduced-functionality designs can create energy-optimization opportunities alongside performance tradeoffs.
5. FUTURE DATABASE DIRECTIONS
Future database research should treat energy efficiency as a system-wide design objective, spanning physical data placement, algorithms, scheduling, resource management, and hardware–software coordination. The paper identifies opportunities ranging from configuration changes to redesigned data structures and cross-layer co-design.
- Physical database design: Physical database design should account for energy use when choosing where and how data is stored.More than half of power use is concentrated in the disk subsystem, motivating renewed attention to storage tradeoffs and physical locations.
- Query optimization and processing: Energy-aware query optimization and processing must revisit assumptions about memory, storage accesses, and CPU requirements.Existing algorithms were designed around performance-oriented resource assumptions that may change under energy objectives.
- Query scheduling and memory management: Scheduling and memory management should be re-examined for energy efficiency, with work sharing across queries becoming increasingly attractive.The traditional focus has been response time and resource utilization rather than energy efficiency.
- Storage and implementations: Energy costs motivate revisions to buffer, storage, logging, recovery, and specialized database implementations.Logging accounts for about 15% of executed code in an online transaction processing system, while specialized systems can reduce unnecessary complexity.
- Co-design: Database systems should coordinate with hardware and other controllers so independent power-management decisions do not work at cross purposes.The paper calls for efficient handoffs, information sharing across controllers, and cooperation spanning architecture, operating systems, and control theory.
- Total cost of ownership: Energy-efficient configurations may require sacrificing hardware cost or parallelizing with additional resources when performance constraints remain.The paper anticipates that rising energy costs will make using more hardware at the same efficiency preferable to wasting energy for diminishing performance returns.
6. CONCLUSIONS
The paper concludes that hardware configurations and application-agnostic power management are only part of the solution to rising data-center energy costs. It argues that data management software has a significant role in improving energy efficiency and urges the community to shift toward energy-efficient computing.
- Data management software will play a significant role in optimizing energy efficiency beyond hardware-only approaches.
- The paper identifies solution approaches and database-system areas that are promising for energy improvements.
- The authors urge database systems researchers to shift focus from performance-oriented research to energy-efficient computing.