- author = "Juan Carlos {Salinas Hilburg}",
- title = "Rapid runtime power and performance profiling of large scale applications",
- school = "Facultad de Informatica, Universidad Complutense de Madrid",
- year = "2021",
- address = "Spain",
- month = "24 " # nov,
- keywords = "genetic algorithms, genetic programming, Grammatical Evolution",
-
URL = "
https://produccioncientifica.ucm.es/documentos/61303681c705c256da91d961",
-
URL = "
https://eprints.ucm.es/id/eprint/67200/1/T42709.pdf",
-
URL = "https://dialnet.unirioja.es/servlet/tesis?codigo=290572",
- size = "152 pages",
- abstract = "Data centres are one of the most power hungry sections of the Information and Communications Technologies (ICT) sector. In the U.S in 2014, data centres consumed around the 1.8percent of the total U.S electricity consumption. Worldwide data centers consumed in 2015 around 200 TWh of the global electricity usage. This electricity consumption is expected to increase to around 1200 TWh in 2025, which would represent 4.5percent of the global electricity usage. One of the major contributors to the overall data centre power is the IT or computing power, therefore there is a special interest to improve its energy efficiency. Scientific community has developed energy efficient techniques to reduce the energy consumption of IT equipment, such as resource management, power budgeting or power capping. These techniques assume the existence of a full dynamic power profiling, obtained through a previous full execution of the application. This full dynamic profiling is not viable in scenarios of long-running applications that are deployed in data centers, since performing a full dynamic profiling of a large batch of long-running applications is a time consuming process thus not energy-efficient. Therefore, we propose the use of an application signature to estimate the energy in a fast way without the need to execute the application from beginning to end. The application signature is a reduced version, in terms of execution time, of the original application. We developed a fast energy estimation framework that uses the application signature to make a quick energy estimation of long-running applications. The framework estimates, without performing a full profile of the application, the dynamic CPU and memory energy of both single-threaded and multi-threaded long-running application versions. Additionally, the fast energy framework is automatic and it has a modular design, allowing to change the functionality of each module without altering the functionality of the whole framework. We validated the accuracy of the fast energy estimation framework with a set of sequential and multi-threaded long-running applications. For the single-threaded version of the applications we obtained an RMS of 10.4percent for the CPU energy estimation error and an RMS of 16.8percent for the memory energy estimation error. In the multi-threaded scenario, we used a subset of applications from the sequential version set. We achieved an RMS of 11.4percent for the CPU energy estimation error and an RMS of 12.8percent for the memory energy estimation error. We defined the concept of Compression Ratio (CR) as the ratio of total execution time of the original application, to the time it takes to estimate the energy through the fast energy estimation framework. A high CR value indicates that the energy is estimated much faster (CR times faster) than executing the whole application. We obtained Compression Ratios in the range from 10.1 to 191.2. Finally, we validated the usefulness of the energy estimation obtained from the application signature by applying three different energy-efficient task scheduling approaches: i) An optimal approach using a Mixed Integer Linear Programming (MILP) technique, ii) An energy-aware heuristic approach that uses a Longest Task First (LTF) algorithm together with an energy-efficient task allocation based on the current servers consumption, and iii) We proposed an implementation of a metaheuristic using a Simulated Annealing process. The results obtained through the energy estimation (obtained through the application signature) values are compared with the real energy values. We obtained energy savings from 8percent to 19percent, and more importantly the energy savings obtained with the application signature approach are similar to the values obtained with the real energy measurements, with an energy savings difference below 1.5percent",
-
notes = "GE (Appendix A) used 'to develop the power (Section
4.2.1.1.2) and co-allocated hardware counters (Section
4.2.2) models'
Supervised by: Jose Luis Ayala Rodrigo and Marina Zapater Sancho and Jose Manuel Moya Fernandez",
