Energy efficiency is a major criterion for computing in general and High Performance Computing in particular. When optimizing for energy efficiency, it is essential to measure the underlying metric: energy consumption. To fully leverage energy measurements, their quality needs to be well-understood. To that end, this thesis provides a rigorous evaluation of various energy measurement techniques. I demonstrate how the deliberate selection of instrumentation points, sensors, and analog processing schemes can enhance the temporal and spatial resolution while preserving a well-known accuracy. Further, I evaluate a scalable energy measurement solution for production HPC systems and address its shortcomings.
Such high-resolution and large-scale measurements present challenges regarding the management of large volumes of generated metric data. I address these challenges with a scalable infrastructure for collecting, storing, and analyzing metric data. With this infrastructure, I also introduce a novel persistent storage scheme for metric time series data, which allows efficient queries for aggregate timelines.
To ensure that it satisfies the demanding requirements for scalable power measurements, I conduct an extensive performance evaluation and describe a productive deployment of the infrastructure.
Finally, I describe different approaches and practical examples of analyses based on energy measurement data. In particular, I focus on the combination of energy measurements and application performance traces. However, interweaving fine-grained power recordings and application events requires accurately synchronized timestamps on both sides. To overcome this obstacle, I develop a resilient and automated technique for time synchronization, which utilizes crosscorrelation of a specifically influenced power measurement signal. Ultimately, this careful combination of sophisticated energy measurements and application performance traces yields a detailed insight into application and system energy efficiency at full-scale HPC systems and down to millisecond-range regions.:1 Introduction
2 Background and Related Work
2.1 Basic Concepts of Energy Measurements
2.1.1 Basics of Metrology
2.1.2 Measuring Voltage, Current, and Power
2.1.3 Measurement Signal Conditioning and Analog-to-Digital Conversion
2.2 Power Measurements for Computing Systems
2.2.1 Measuring Compute Nodes using External Power Meters
2.2.2 Custom Solutions for Measuring Compute Node Power
2.2.3 Measurement Solutions of System Integrators
2.2.4 CPU Energy Counters
2.2.5 Using Models to Determine Energy Consumption
2.3 Processing of Power Measurement Data
2.3.1 Time Series Databases
2.3.2 Data Center Monitoring Systems
2.4 Influences on the Energy Consumption of Computing Systems
2.4.1 Processor Power Consumption Breakdown
2.4.2 Energy-Efficient Hardware Configuration
2.5 HPC Performance and Energy Analysis
2.5.1 Performance Analysis Techniques
2.5.2 HPC Performance Analysis Tools
2.5.3 Combining Application and Power Measurements
2.6 Conclusion
3 Evaluating and Improving Energy Measurements
3.1 Description of the Systems Under Test
3.2 Instrumentation Points and Measurement Sensors
3.2.1 Analog Measurement at Voltage Regulators
3.2.2 Instrumentation with Hall Effect Transducers
3.2.3 Modular Instrumentation of DC Consumers
3.2.4 Optimal Wiring for Shunt-Based Measurements
3.2.5 Node-Level Instrumentation for HPC Systems
3.3 Analog Signal Conditioning and Analog-to-Digital Conversion
3.3.1 Signal Amplification
3.3.2 Analog Filtering and Analog-To-Digital Conversion
3.3.3 Integrated Solutions for High-Resolution Measurement
3.4 Accuracy Evaluation and Calibration
3.4.1 Synthetic Workloads for Evaluating Power Measurements
3.4.2 Improving and Evaluating the Accuracy of a Single-Node Measuring System
3.4.3 Absolute Accuracy Evaluation of a Many-Node Measuring System
3.5 Evaluating Temporal Granularity and Energy Correctness
3.5.1 Measurement Signal Bandwidth at Different Instrumentation Points
3.5.2 Retaining Energy Correctness During Digital Processing
3.6 Evaluating CPU Energy Counters
3.6.1 Energy Readouts with RAPL
3.6.2 Methodology
3.6.3 RAPL on Intel Sandy Bridge-EP
3.6.4 RAPL on Intel Haswell-EP and Skylake-SP
3.7 Conclusion
4 A Scalable Infrastructure for Processing Power Measurement Data
4.1 Requirements for Power Measurement Data Processing
4.2 Concepts and Implementation of Measurement Data Management
4.2.1 Message-Based Communication between Agents
4.2.2 Protocols
4.2.3 Application Programming Interfaces
4.2.4 Efficient Metric Time Series Storage and Retrieval
4.2.5 Hierarchical Timeline Aggregation
4.3 Performance Evaluation
4.3.1 Benchmark Hardware Specifications
4.3.2 Throughput in Symmetric Configuration with Replication
4.3.3 Throughput with Many Data Sources and Single Consumers
4.3.4 Temporary Storage in Message Queues
4.3.5 Persistent Metric Time Series Request Performance
4.3.6 Performance Comparison with Contemporary Time Series Storage Solutions
4.3.7 Practical Usage of MetricQ
4.4 Conclusion
5 Energy Efficiency Analysis
5.1 General Energy Efficiency Analysis Scenarios
5.1.1 Live Visualization of Power Measurements
5.1.2 Visualization of Long-Term Measurements
5.1.3 Integration in Application Performance Traces
5.1.4 Graphical Analysis of Application Power Traces
5.2 Correlating Power Measurements with Application Events
5.2.1 Challenges for Time Synchronization of Power Measurements
5.2.2 Reliable Automatic Time Synchronization with Correlation Sequences
5.2.3 Creating a Correlation Signal on a Power Measurement Channel
5.2.4 Processing the Correlation Signal and Measured Power Values
5.2.5 Common Oversampling of the Correlation Signals at Different Rates
5.2.6 Evaluation of Correlation and Time Synchronization
5.3 Use Cases for Application Power Traces
5.3.1 Analyzing Complex Power Anomalies
5.3.2 Quantifying C-State Transitions
5.3.3 Measuring the Dynamic Power Consumption of HPC Applications
5.4 Conclusion
6 Summary and Outlook
| Identifer | oai:union.ndltd.org:DRESDEN/oai:qucosa:de:qucosa:71600 |
| Date | 31 July 2020 |
| Creators | Ilsche, Thomas |
| Contributors | Nagel, Wolfgang E., Schulz, Martin, Technische Universität Dresden |
| Source Sets | Hochschulschriftenserver (HSSS) der SLUB Dresden |
| Language | English |
| Detected Language | English |
| Type | info:eu-repo/semantics/publishedVersion, doc-type:doctoralThesis, info:eu-repo/semantics/doctoralThesis, doc-type:Text |
| Rights | info:eu-repo/semantics/openAccess |
| Relation | info:eu-repo/grantAgreement/Deutsche Forschungsgemeinschaft/Sonderforschungsbereich/SFB 912//Highly Adaptive Energy-efficient Computing /HAEC |
Page generated in 0.0018 seconds