MSN 数学媒体与信息存储 1/27 Zhuo Liu, Fei Wu, Xiao Qin, Changsheng Xie, Jian Zhou, and Jianzong Wang TRACER: A Trace Replay Tool to Evaluate Energy-Efficiency of Mass Storage Systems WNLO, Huazhong University of Science and Technology; Computer Science Department, Auburn University
MSN 数学媒体与信息存储 2/27 Outline I. Motivation Why do we need TRACER? II. Architecture What does TRACER consist of? How does TRACER work and look like? III. Trace-replay control scheme How does TRACER control load intensity? IV. Measurement How well does TRACER work? V. Conclusion and future work
MSN 数学媒体与信息存储 3/27 I Motivation New energy conservation techniques in storage systems constantly spring up: such as MAID, DRPM, PDC… Lack of systematic and uniform way to evaluate them: present benchmarks and standards are most for CPU-intensive applications
MSN 数学媒体与信息存储 4/27 Current power-aware techniques
MSN 数学媒体与信息存储 5/27 Energy-efficiency Standards SPEC (CPU-intensive) Metric: Java_ops/Watt SUN swap (CPU-intensive) Metric: Performance/(Space*Watt) Joule Sort (CPU-intensive) Metric: Sort_ops/Joule Energy Star SNIA green Classification of storage devices Standards mainly for idle-mode tests
MSN 数学媒体与信息存储 6/27 Power increases with load intensity SNIA-Green
MSN 数学媒体与信息存储 7/27 What should TRACER do? Test the power consumptions Produce different modes of IO load Regulate IO load intensity: 10%-100% Metrics: to evaluate energy-efficiency TRACER
MSN 数学媒体与信息存储 8/27 II. Architecture of TRACER TRACER consists of four parts: Evaluation host Workload generator Power analyzer Storage systems under test
MSN 数学媒体与信息存储 9/27 What does TRACER consist of?
MSN 数学媒体与信息存储 10/27 How is TRACER implemented?
MSN 数学媒体与信息存储 11/27 GUI of TRACER
MSN 数学媒体与信息存储 12/27 III. Trace-replay control scheme The structure of a blocktrace file
MSN 数学媒体与信息存储 13/27 Load-control algorithm for trace replay
MSN 数学媒体与信息存储 14/27 IV. Measurement: environment
MSN 数学媒体与信息存储 15/27 Measurement: hardware connects
MSN 数学媒体与信息存储 16/27 Measurement : traces Three types of traces: 125 traces we connected using IOmeter Request size: 512B, 4KB, 16KB, 64KB, 1MB Random/sequential%: 0%,25%, 50%, 75%,100% Read/write%: 0%, 25%,50%,75%, 100% HP lab: cello 96 and cello 99 The web server trace (FAST’09-BORG)
MSN 数学媒体与信息存储 17/27 Measurement : evaluation metrics IO Throughput/Power Consumption IOPS/Watt MBPS/Kilowatt
MSN 数学媒体与信息存储 18/27 V. Measurement: results 1. Accuracy of load control for 125 traces Accuracy % =Measured intensity% / configured% 2. How Energy-efficiency is influenced By load intensity By random% By read% 3. Accuracy for Real traces HP lab cello99 cello96, web server traces
MSN 数学媒体与信息存储 19/27 Accuracy of load control
MSN 数学媒体与信息存储 20/27 E-E influenced by load%
MSN 数学媒体与信息存储 21/27 E-E influenced by random%
MSN 数学媒体与信息存储 22/27 E-E influenced by read%
MSN 数学媒体与信息存储 23/27 Load control for real trace
MSN 数学媒体与信息存储 24/27 Accuracy of load control for real trace
MSN 数学媒体与信息存储 25/27 Conclusion and Future Work 1 TRACER is accurate, efficient and useful. 2 Storage system is more energy-efficient under higher load intensity, under lower random rate. 3 As temperature influences system’s performance and power, it’s necessary to add temperature as part of evaluation metrics.
MSN 数学媒体与信息存储 26/27 Add Temperature metrics Temp sensor
MSN 数学媒体与信息存储 27/27 Any question?