Self-Adaptive Power Management of Idle Nodes in Large Scale Systems

Abstract

The real workload on a large scale computer system varies from time to time. Often it has many idle nodes during most operation time. These idle nodes consume energy, but do nothing useful. To save the huge amount of energy wasted by such active idle nodes, most modern compute nodes are equipped with multiple level dynamic sleep mechanisms to reduce power consumption. However, awaking sleeping nodes takes time. The deeper a node sleeps, the less energy it consumes, but the longer wakeup latency. How to balance between the systems energy consumption and the response time is a key problem in the power management of large scale systems. This paper proposes a self-adaptive approach to manage the sleep states of idle nodes to achieve low energy consumption and high performance at the same time. The proposed approach has two distinctive features. First, idle nodes are hierarchical organised. In this model, idle nodes are classified into several groups according to their sleep states. Each group contains nodes of same level of sleep depth and forms a reserve pool of a certain readiness level. When a resource is requested, nodes in the pool of highest level of readiness are preferentially allocated. When the nodes in the pool of the highest readiness level are not sufficient, the nodes in the pool(s) of next level(s) of readiness are allocated. After each allocation and reclaim of nodes, the numbers of nodes in each level of pools are adjusted by changing the sleep depth of the nodes up and down. Thus, the reserve pools can be maintained for high performance requirement. When resources are released from applications, they are placed back to reserve pools and put into different levels of sleep states to save energy. Second, the sizes of reservation pools are self-adaptive. Obviously, a key factor that affects the effectiveness of the idle node management is the sizes of the reserve pools. Fixed sizes of reserve pools would not be effective due to the time varying nature of workload on large scale systems. The proposed approach employs a self-adaptive mechanism in which the sizes of reserve pools are dynamically adjusted during the execution of the system according to how well the research pools meet the need of computation resources. Our experiments demonstrated that our approach can significantly improve energy efficiency in large scale systems without significant scarification of performance.