There are various techniques for reducing power consumption of service platforms on a cluster system. However, such techniques are for managing the power of hardware resources within a single service platform. Therefore, it is difficult to support eff...
There are various techniques for reducing power consumption of service platforms on a cluster system. However, such techniques are for managing the power of hardware resources within a single service platform. Therefore, it is difficult to support efficient power management for a cluster system in a data center which is widely used by internet portals.
Power management mechanisms can affect other aspects of system behaviors such as performance and availability. For example, in case of performing power control of a single service platform using DVFS, performance of some workloads degrades proportional to CPU frequency reduction, while other workloads are largely unaffected. In case of using ACPI, service platforms constituting the cluster system often go down or go to sleep. Therefore, it frequently occurs that services are not smoothly provided when there are many service requests. Moreover, as the complexity of a monitoring system increases, the cost for managing the system becomes high. For these reasons, power cannot be managed independently. In this dissertation, we are concerned with a policy based power management scheme for energy efficient cluster systems.
The main contributions of this dissertation are summarized as follows. First, we survey the function and the method of the policy based power management which has been applied to many data centers and cluster systems. Also, the power control and platform management techniques for service platforms in order to configure a cluster system are surveyed.
Second, we propose the policy based power management techniques which are capable of reducing power consumption in an entire cluster system. The action state is controlled based on the utilization of service platforms constituting a cluster system without degrading of the performance. To validate the proposed policy-based dynamic power management techniques, we implement them on an actual cluster system for web-services, where service platforms support all the power management techniques. Also, we compare and analyze the power reduction with the policy-based power management and without the cluster system.
Third, we propose the power modeling of service platform which is used to predict the power consumption of a service platform. We compare the measured power consumption and the predicted power consumption of a service platform during an ftp service. Using the comparison results, we analyze the accuracy of the power modeling and determine whether it can forecast the total power consumption of a cluster system or not.
Forth, we propose the optimal threshold of working speed which is used to determine the CPU frequency for service platform level power management. To determine the optimal threshold of working speed, we compare the power consumption of a service platform while changing the threshold. Using the comparison results, we suggest the optimal DPM(Dynamic Power Management) techniques and the threshold of working speed for a high-performance, energy efficient cluster system.
Power consumption is now a major area of concern for designing a high performance cluster system. The dynamic power management techniques for reducing the power consumption based on a single service platform are considered to achieve not only for high performance but for energy efficiency in a cluster system. In selecting the policy based dynamic power management technique for the design of a high-performance and energy efficient cluster system, we expect that our result in this dissertation becomes a useful criterion.