A Game Theoretic Approach to Power Reduction in Distributed Storage Systems
A Game Theoretic Approach to Power Reduction in Distributed Storage Systems
复制标题
DOI：
10.2197/ipsjjip.24.173
复制
发表时间：
2016
期刊：
影响因子：
--
通讯作者：
Koji Hasebe;Takumi Sawada;Kazuhiko Kato
中科院分区：
文献类型：
--
作者：
Koji Hasebe;Takumi Sawada;Kazuhiko Kato
本文提出了一种大规模分布式存储系统功耗降低的博弈论方法。关键思想是使用分布式哈希表并动态迁移其虚拟节点，以便将工作负载倾斜到物理磁盘子集，同时不会使它们过载。为了以一种自治的方式实现这个想法，虚拟节点被视为自私的代理，在这个游戏中，每个节点根据其当前所在磁盘的工作负载获得收益。我们将此设置建模为潜在博弈，即所有玩家改变策略的动机可以用单一全局函数表示的一种战略博弈。因此，虚拟节点收益的任何增加都会在能量节约方面产生更好的状态。该博弈模型由一对全局和私有收益函数组成，由美好生活效用方案导出。前一个函数评估系统的当前状态有多好，而后一个函数确定每个节点的当前收益。通过仿真和原型实现对该方法的性能进行了测试。实验结果表明，该方法比静态结构能耗降低11.1% ~ 16.4%。此外，尽管少数响应由于某些磁盘在峰值时间过载而严重延迟，但我们的方法在50-190 ms范围内保持了首选的总体平均响应时间。
We present a game theoretic approach for power reduction in large-scale distributed storage systems. The key idea is to use a distributed hash table and migrate its virtual nodes dynamically so as to skew the workload towards a subset of physical disks while not overloading them. To realize this idea in an autonomous way, virtual nodes are regarded as selfish agents playing a game in which each node receives a payoff according to the workload of the disk on which it currently resides. We model this setting as a potential game, a kind of strategic game in which the incentive of all players to change their strategy can be represented by a single global function. Thus, any increase in the payoff of a virtual node yields a better state in terms of energy conservation. This game model consists of a pair of global and private payoff functions, derived by the Wonderful Life Utility scheme. The former function evaluates how good the current state of the system is, while the latter determines the current payoff of each node. The performance of our method is measured both by simulations and a prototype implementation. From the experiments, we observed that our method consumed 11.1%–16.4% less energy than the static configuration. In addition, although a small number of responses were heavily delayed because of overloading of some disks at peak time, our method maintained preferred overall average response time in the range 50–190 ms.