Improving Multicore Server Performance and Reducing Energy Consumption by Workload Dependent Dynamic Power Management

• Published in: IEEE transactions on cloud computing Vol. 4; no. 2; pp. 122 - 137
• Main Author: Li, Keqin
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE Computer Society 01.04.2016; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Cloud computing; Dynamic power management; Dynamics; Energy consumption; Microprocessors; Multicore processing; multicore server processor; Power consumption; Power demand; Power management; Power supply; Product development; queueing model; Response time; Servers; Tasks; Time factors; Workload; Dynamic power management; multicore server processor; queueing model; response time; energy consumption
• ISSN: 2168-7161; 2372-0018
• DOI: 10.1109/TCC.2015.2440238

The technique of using workload dependent dynamic power management (i.e., variable power and speed of processor cores according to the current workload) to improve system performance and to reduce energy consumption is investigated. Typically, the power supply and the core speed are increased when there are more tasks in a server, such that tasks can be processed faster and the average task response time is reduced. On the other hand, the power supply and the core speed are decreased when there are less tasks in a server, such that energy consumption can be reduced without significant performance degradation. A queueing model of multicore server processors with workload dependent dynamic power management is established. Several speed schemes are proposed and it is demonstrated that for the same average power consumption, it is possible to design a multicore server processor with workload dependent dynamic power management, such that its average task response time is shorter than a multicore server processor of constant speed (i.e., without workload dependent dynamic power management). It is shown that given certain application environment and average power consumption, there is an optimal speed scheme that minimizes the average task response time. For two-speed schemes, the problem of optimal design of a two-speed scheme for given power supply and power consumption model is formulated and solved. It is pointed out that power consumption reduction subject to performance constraints can be studied in a similar way as performance improvement (i.e., average task response time reduction) subject to power consumption constraints. To the best of our knowledge, this is the first work on analytical study of workload dependent dynamic power management.