基于资源配置等效性的数据中心能耗优化

Improvement of datacenters' energy efficiency based on resource allocation equivalency

针对数据中心服务器的低能效问题,进行了利用资源配置的等效性来优化服务器能效比的研究。研究发现,应用程序的多种资源分配方案具有相同的性能,但表现出较大的能耗差异,这种现象叫做"基于性能等效的资源配置",简称"等效配置"。基于这种观察,提出了两种优化能效比的算法——SmartRank算法和Smart Balance算法。SmartRank算法使用资源等效替换的方法寻找能耗最低的资源配置,来达到局部最优的能效比;Smart Balance算法通过评估资源需求向量与剩余资源间的关系来均衡资源分配,同时兼顾单个应用的能耗开销,从而达到全局最大能效比。实验表明,通过对这两个算法的优化,可实现平均节省3%的系统能耗,局部最大可以节省12.5%的能耗。

The equivalence of resource allocation was used to improve the energy-efficiency proportion of servers to solve the low energy efficiency problem of modern datacenters. The study observed that variety of resource allocation schemes for applied programs have the same performance but they are apparently different in energy consumption. This phenomenon was called the equivalent allocation. Based on the observation, two algorithms for improving the energy-efficiency proportion, named SmartRank and SmartBalance, were proposed. The SmartRank algorithm uses the resource equivalent replacement to seek the resource allocation consuming lowest energy to achieve the energy- efficiency proportion of local optimization. The SmartBalance algorithm balances the resource allocation through evaluation of the relation between the vector of resource demand and the surplus resources while takes account of the cost for energy consumption to achieve the maximum local energy-efficiency proportion. The experiments show that the SmartRank algorithm can reduce the energy of a single application as much as 12.5%, and the SmartBal- ance algorithm can save 3% of energy of the whole system on average.