Citation Context

...ce is still an open challenge. However, there is an increasing interest in energy-proportional servers [6] and dynamic server provisioning techniques are being explored by both academics and industry =-=[7, 8, 9, 10, 11]-=-. Results. To conduct our analysis, we use trace-driven simulation with real-world hourly (and daily) energy prices obtained from a number of data sources. We look at 39 months of hourly electricity p...

Citation Context

...irements DVFS Coda and Odyssey, Flinn and Satyanarayanan [33] CPU, network interface Mobile systems Minimize energy consumption by application degradation Resource throttling PowerNap, Meisner et al. =-=[34]-=- System-wide Server systems Minimize power consumption, minimize performance loss DCD Application adaptation System resources No Adapted applications Single resource Multiple resources Arbitrary Targe...

Citation Context

...on (hence, achieved energy saving). An increase in the number of supported rates is also shown to result in better performance, at the price of an increased management complexity. Other works such as =-=[46]-=-, [47] provide a relative comparison of sleeping mode versus rate switch strategies, when applied to processors and servers, respectively. In this case, both works are in favor of sleeping mode strate...

Citation Context

...th. By 2011, U.S. data centers are predicted to use 100 Billion kWh, at a cost of $7.4 billion per year [50]. Of particular concern is the low average data center utilization, typically around 20-30% =-=[10, 37]-=-, which when coupled with high idle power consumption (at idle, current servers use about 60% of peak power), leads to significant waste. The combination of brief but frequent bursts of activity coupl...

Citation Context

...ly, the third class of energy management techniques consider opportunistically duty-cycling subsystems, such as wireless radios[2, 22, 24], networking infrastructure [14, 21] or even entire platforms=-=[3, 18, 25]-=-, during periods of idleness or low use. SleepServer falls into this third category of energy management approaches. The duty-cycling technique exploits the capability of modern hardware to enter low-...

Citation Context

...In contrast to this work we operate at a granularity of seconds, giving us more agility in dealing with sharp changes in request rate. Several papers have proposed putting machines to sleep. PowerNap =-=[17]-=- models how quickly servers would have to transition in and out of sleep states to achieve energy proportionality on web and other workloads. The authors find that a transition time of under 10ms is r...

Citation Context

... mode after approximately 32 minutes of activity. This value is effective for real-world systems, as modern servers allow low-latency transitions to the sleep mode consuming low power. Meisner et al. =-=[42]-=- have shown that a typical blade server consuming 450 W in the fully utilized state consumes approximately 10.4 W in the sleep mode, while the transition delay is 300 ms. The mean number of host trans...

Citation Context

... not concentrate the loading on fewer servers, temporal subsetting may not incur the temperature problems of spatial subsetting. However, DVFS transition times are too long and degrade response times =-=[20]-=-, which is the second issue in reducing idle power (relevant to both spatial subsetting and temporal subsetting). In addition, DVFS is applicable only to the CPU, which contributes only about one-thir...

Citation Context

..., 4, 1]. This paper argues that datacenter management should be revised to maximize the use of such off-grid renewables. 1 Our argument is similar in spirit to the case for energyefficient management =-=[18, 7, 11]-=-. But efficiency alone may not take full advantage of renewables. For instance, energy-efficient management may batch writes so disks can sleep for long durations. But sometimes renewables are intermi...