Modelagem, caracterização e otimização de potência em centro de dados

Author(s)

Leonardo de Paula Rosa Piga

• Document type: Doctoral Thesis
• Press: Campinas, SP.
• Institution: Universidade Estadual de Campinas (UNICAMP). Instituto de Computação
• Defense date: 2013-08-11

Examining board members

Sandro Rigo; Philippe Olivier Alexandre Navaux; Hermes Senger; Edmundo Roberto Mauro Madeira; Rodolfo Jardim de Azevedo

• Advisor: Sandro Rigo; Reinaldo Alvarenga Bergamaschi

Abstract

To keep up with an increasing demand for computational resources, IT companies need to build facilities that host hundreds of thousands of computers, the data centers. This environment is highly dependent on electrical energy, a resource that is becoming expensive and limited. In this context, this thesis develops a global data center-level power and performance optimization approach for Web Server data centers. It presents a power measurement framework for commodity servers, develops empirical models for estimating the power consumed by Web servers, and implements one of the global power optimization heuristics on a state-of-the-art, high-density SeaMicro SM15k cluster by AMD. The power measuring framework is composed of a custom made board, which is able to capture the power consumption; a data acquisition device that samples the measured values; and a piece of software that manages the framework. We show a novel method for developing full system Web server power models that prunes model parameters and reduces non-linear relationships among performance measurements and system power. The Web server power models use as parameters performance indicators read from the machine internal performance counters. We evaluate our approach on an AMD Opteron-based Web server and on an Intel i7-based Web server. Our best model displays an average absolute error of 1.92% for the Intel i7 server and 1.46% for AMD Opteron as compared to actual measurements, and 90th percentile for the absolute percent error equals to 2.66% for Intel i7 and 2.08% for AMD Opteron. We deploy the global power management system in a state-of-the-art SeaMicro SM15k cluster. The implementation relies on the concept of Virtual Power States, a combination of CPU utilization rate to the P/C power states available in modern processors, and on our global optimization algorithm called Slack Recovery. We also propose and implement a novel mechanism to control utilization rates in each server, a key aspect of our power/performance optimization system. Experimental results show that our Slack Recovery-based system can reduce up to 16% of the power consumption when compared to the Linux performance governor and 6.7% when compared to the Linux ondemand governor (AU)

• FAPESP's process: 10/05389-5 - Power and performance modelling and analysis of datacenter servers
• Grantee: Leonardo de Paula Rosa Piga
• Support Opportunities: Scholarships in Brazil - Doctorate (Direct)