Response-Time Modeling for Resource Allocation and Energy-Informed SLAs
Peter Bodík, Charles A. Sutton, Armando Fox, David A. Patterson, Michael I. Jordan · Edinburgh Research Explorer (University of Edinburgh) · 2007
The growth of Internet datacenters is increasingly limited by their power consumption. Powering and cooling a datacenter now rivals the cost of the hardware: each $1 spent on servers in 2005 required an additional $0.48 to power and cool it, expected to rise to $0.71 by 2010 [4]. Techniques to scale down power consumption during periods of low workload are therefore desirable. However, datacenter operators also typically require a strong guarantee on the tail of the response-time distribution called a service-level agreement (SLA), such as that the 99th percentile of response times is less than 300 ms. Power management technology built in to current microprocessors includes techniques such as dynamic voltage and frequency scaling (DVFS), which dynamically reduces the CPU frequency during periods of low utilization. In actual datacenters, however, the built-in DVFS is almost never used, because conventional wisdom holds that its use leads to unpredictable service times. Indeed, this concern is justified, because built-in power management strategies are oblivious to the desired SLA, so they cannot increase resources when the desired performance objective is not met.