Hydra:高能效的异构服务器融合方法

Hydra:Improving Energy Efficiency Through Multiple Heterogeneous Servers

考虑当前数据中心服务器面临日常工作负载强度从0%~90%的大幅波动,但很多时间服务器处于低利用率或者空闲状态.即使在空闲状态,典型服务器仍然需要消耗50%左右峰值电力来维持其活动状态.然而简单地关闭空闲服务器的策略很难实施,因为不可预测的用户请求和突发事件可能导致频繁开关机,进而产生严重的性能和可靠性降级.提出了一种新颖异构服务器级融合的节能解决方案Hydra.Hydra提出了一种新架构,该架构把多台异构(功率/性能/成本)特性的服务器整合成虚拟服务器,并根据负载变化进行自动切换.为此Hydra提供了2个关键机制:1)Hydra监听机制,监控传入的负载强度来决定何时进行服务器切换;2)Hydra切换机制,能够融合各服务器的空闲和忙碌时间段,为系统级深度睡眠创造机会;Hydra是基于2个概念设计的:1)在轻负载条件下,优化能源利用效率;2)保证响应时间上界.最后实验证明,Hydra降低服务器70%的平均能耗.同时通过成本模型,发现Hydra能够提升18%的每TCO(total cost of ownership)美元的性能.

Considering servers in current data centers endure significant workload fluctuation from nearly 0%to 90% over time,but the most of the time,servers are underutilized or idle.Even if the servers are in the idle state,about 50% peak power has been wasted to maintain typical servers to be active even under light loads or idle.However simply turning a idle system off cannot achieve the energy efficiency,the unpredictable and burst user requests would result in an intolerable performance and reliability degradation.In the paper,We propose Hydra,a novel energy-saving solution for heterogeneous servers.Hydra presents a new architecture,it can consolidate hybrid power/performance/cost nodes as a virtual server,and switch automatically as the load changes.Hydra also consists of two critical mechanisms:1)Hydra-listener,a monitor that detects incoming task traffic to determine when a power transition should occur;2)Hydra Transition,a scheduling policy to coalesce idle and busy periods between engines to create opportunities for the system-level deep sleep.Hydra is based on two principles:1)optimize energy efficiency while load is low,and 2)guarantee a bounded response time.Finally,our experiments demonstrate that Hydra reduces average server power consumption by 70% and improves performance per TCO(total cost of ownership)dollar by 18%.