Optimal Power Allocation in Server Farms服务器场的最优功率分配.pptVIP

Optimal Power Allocation in Server Farms Anshul Gandhi(CMU), Mor Harchol-Balter(CMU), Rajashri Das(IBM T.J. Watson),Charles Lefurgy(IBM Austin) Motivation Cost of energy is rising a 300W high performance server requires more than $330 of energy cost per year Server farms usually have a fixed peak power budget Are billed based on their peak power requirements Peak power budges also determines the cooling and power delivery infrastructure costs Why is this problem difficult ? No single power allocation is optimal in all scenarios Optimal power allocation depends on a lot of factors Arrival rate, open or closed workload, power-to-frequency curve, maximum power a server can use, minimum power a server can use Contributions Experimental Setup 14 IBM BladeCenter blade servers in a single chassis 3 GHz dual-core Intel processors 1 GB memory per blade Apache installed on each blade Another blade server on same chassis is workload generator httperf, open configuration wbox, closed configuration Speed scaling Reduce power by reducing server speed Intel’s SpeedStep , AMD’s Cool ‘n’ Quiet Speed Scaling mechanisms DFS 8 operating points 12.5%,25%,37.5%,50%…..100% of maximum server frequency DVFS 4 operating points 66.5%,77.7%…100% DVFS + DFS 11 operating points Apply DFS on the lowest performance state available in DVFS Power consumption of a server Baseline power: Pmin watts Minimum power consumed by a fully-utilized server over the entire range of frequencies The speed at which it consumes this power smin Maximum power: Pmax Watts Maximum power consumed by a fully utilized server over the entire range of frequencies The speed at which it consumed this power smax These values depend on the workload Terms Power-to-Frequency curve Apply a particular power cap on a server Run jobs back-to-back to ensure the server is fully utilized Shift between frequency states to ensure you don’t exceed the maximum power System-level power, NOT just CPU Power-to-frequency curve We a