T

Aa

b

c

ARA

KPAF

1

fnttt(cpiaegsai

(((

2h

Sustainable Computing: Informatics and Systems 3 (2013) 183– 193

Contents lists available at SciVerse ScienceDirect

Sustainable Computing: Informatics and Systems

j ourna l h om epage: www.elsev ier .com/ locate /suscom

he need for speed and stability in data center power capping

rka A. Bhattacharyaa,∗, David Cullera, Aman Kansalb, Sriram Govindanc, Sriram Sankarc

University of California, Berkeley, Berkeley, CA, USAMicrosoft Research, Redmond, WA, USAMicrosoft Corporation, Redmond, WA, USA

a r t i c l e i n f o

rticle history:eceived 11 September 2012ccepted 29 January 2013

eywords:ower cappingdmission controlrequency scaling

a b s t r a c t

Data centers can lower costs significantly by provisioning expensive electrical equipment (such as UPS,diesel generators, and cooling capacity) for the actual peak power consumption rather than server name-plate power ratings. However, it is possible that this under-provisioned power level is exceeded due tosoftware behaviors on rare occasions and could cause the entire data center infrastructure to breach thesafety limits. A mechanism to cap servers to stay within the provisioned budget is needed, and processorfrequency scaling based power capping methods are readily available for this purpose. We show thatexisting methods, when applied across a large number of servers, are not fast enough to operate cor-rectly under rapid power dynamics observed in data centers. We also show that existing methods whenapplied to an open system (where demand is independent of service rate) can cause cascading failures

in the software service hosted, causing the service performance to fall uncontrollably even when powercapping is applied for only a small reduction in power consumption. We discuss the causes for both theseshort-comings and point out techniques that can yield a safe, fast, and stable power capping solution. Ourtechniques use admission control to limit power consumption and ensure stability, resulting in orders ofmagnitude improvement in performance. We also discuss why admission control cannot replace existingpower capping methods but must be combined with them.

. Introduction

The cost of provisioning power in data centers is a very largeraction of the total cost of operating a data center [1–3] ranking justext to the cost of the servers themselves. Provisioning costs includehe cost of infrastructure for sourcing, distribution and backup forhe peak power capacity (measured in $/kW). These are higher thanhe consumption costs paid per unit of energy actually consumedmeasured in $/kWh) over the life of a data center. Provisionedapacity and related costs can be reduced by minimizing the peakower drawn by the data center. A lower capacity saves on expenses

n utility connection charges, diesel generators, backup batteries,nd power distribution infrastructure within the data center. Low-ring capacity demands is also greener because from the powereneration standpoint, the cost and environmental impact for large

cale power generation plants such as hydro-electric plants as wells green energy installations such as solar or wind farms, is dom-nated by the capacity of the plant rather than the actual energy

∗ Corresponding author.E-mail addresses: arka@eecs.berkeley.edu, arkaaloke@gmail.com

A.A. Bhattacharya), culler@eecs.berkeley.edu (D. Culler), kansal@microsoft.comA. Kansal), srgovin@microsoft.com (S. Govindan), sriram.sankar@microsoft.comS. Sankar).

210-5379/$ – see front matter © 2013 Elsevier Inc. All rights reserved.ttp://dx.doi.org/10.1016/j.suscom.2013.01.005

© 2013 Elsevier Inc. All rights reserved.

produced. From the utility company perspective, providing peakcapacity is expensive due to the operation of ‘peaker power plants’which are significantly more expensive to operate and are lessenvironmentally friendly than the base plants. Aside from costs,capacity is now is short supply in dense urban areas, and utili-ties have started refusing to issue connections to new data centerslocated in such regions. Reducing the peak power capacity requiredis hence extremely important.

The need to manage peak power is well understood and mostservers ship with mechanisms for power capping [4,5] that allowlimiting the peak consumption to a set threshold. Further capacitywaste can be avoided by coordinating the caps across multipleservers. For instance, when servers in one cluster or applicationare running at lower load, the power left unused could be usedby other servers to operate at high power levels than would beallowed by their static cap. Rather than forcing a lower aggregatepower level at all times, methods that coordinate the power capsdynamically across multiple servers and applications have beendeveloped [6–10].

We identify two reasons why existing power capping methodsdo not adequately meet the challenge of power capping in data

centers. The first is speed. We show through real world data cen-ter power traces that power demand can change at a rate thatis too fast for the existing methods. The second is stability. Weexperimentally show that when hosting online applications, the

1 ting:

sicetova

•

•

•

•

2

tIIDtdd

2

btc

sgpWstiwicotasfd

I$i

84 A.A. Bhattacharya et al. / Sustainable Compu

ystem may become unstable if power capped. A small reductionn power achieved through existing power capping methods canause the application latency to increase uncontrollably and mayven reduce throughput to zero. We focus on the importance of thewo necessary properties – speed and stability, and propose waysf achieving them and discuss the tradeoffs involved. Our obser-ations are generic, and can be integrated into any power cappinglgorithm.

Specifically, the paper makes the following contributions:

We quantify the benefit of using power capping to lower powerprovisioning costs in data centers through the analysis of a realworld data center power trace.Speed requirement: From the same trace, we characterize the ratesat which power changes in a data center. We make a case forone-step power controllers by showing that existing closed-looptechniques for coordinated power capping across a large numberof servers may not be fast enough to handle data center powerdynamics.Stability requirement: We show that existing power capping tech-niques do not explicitly shape demand, and can lead to instabilityand unexpected failures in online applications.We present admission control as a power capping knob. Wedemonstrate that admission control integrated with existingpower capping techniques can achieve desirable stability char-acteristics, and evaluate the trade-offs involved.

. Power costs and capping potential

Most new servers ship with power capping mechanisms. Sys-em management software, such as Windows Power Budgetingnfrastructure, IBM Systems Director Active Energy Manager, HPnsight Control Power Management v.2.0, Intel Node Manager, andell OpenManage Server Administrator, provide APIs and utilities

o take advantage of the capping mechanisms. In this section weiscuss why power capping has become a significant feature forata centers.

.1. Power provisioning costs

The designed peak power consumption of a data center impactsoth the capital expense of provisioning that capacity as well ashe operating expense of paying for the peak since there is often aharge for peak usage in addition to that for energy consumed.

The capital expense (cap-ex) includes power distribution infra-tructure as well as the cooling infrastructure to pump out the heatenerated from that power, both of which depend directly on theeak capacity provisioned. The cap-ex varies from $10 to $25 peratt of power provisioned [3]. For example, a 10 MW data center

pends about $100–250 million in power and cooling infrastruc-ure. Since the power infrastructure lasts longer than the servers,n order to compare this cost as a fraction of the data center expense,

e can normalize all costs over the respective lifespans. Amortiz-ng cap-ex over the life of the data center (12–15 years [3,2]), serverosts over the typical server refresh cycles (3–4 years), and otherperating expenses at the rates paid, the cap-ex is over a third ofhe overall data center expenses [11,2]. This huge cost is primarilyttributable to the expensive high-wattage electrical equipment,uch as UPS batteries, diesel generators, and transformers, and isurther exacerbated by the redundancy requirement mandated byata center availability stipulations.

The peak power use affects operating expenses (op-ex) as well.n addition to paying a per unit energy cost (typically quoted in/kWh), there is an additional fee for the peak capacity drawn, evenf that peak is used extremely rarely. Based on current utility tariffs

Informatics and Systems 3 (2013) 183– 193

[12] for both average and peak power, the peak consumption cancontribute to as much as 40% of the utility bill [13]. Utility com-panies may also impose severe financial penalties for exceedingcontracted peak power limits.

The key implication is that reducing the peak capacity requiredfor a data center, and adhering to it, is highly beneficial.

2.2. Lower cost through capping

Power capping can help manage peak power capacity in severalways. We describe some of the most common reasons to use itbelow.

2.2.1. Provisioning lower than observed peakProbably the most widely deployed use case for power capping

is to ensure safety when power is provisioned for the actual datacenter power consumption rather than based on server nameplateratings. Nameplate ratings on servers denotes its maximum possi-ble power consumption, computed as the sum of maximum powerconsumption of all the server sub-components and a conserva-tive safety margin. The name-plate rating on servers is typicallymuch higher than the server’s actual consumption. Since no work-load actually exercises every server subcomponent at its peak ratedpower, the name plate power is not reached in practice. Data cen-ter designers thus provision for the observed peak on every server.The observed peak is the maximum power consumption measuredon a server when running the hosted application at the highestrequest rate supported by the server. This observed peak can beexceeded after deployment due to software changes or events suchas server reboots that may consume more than the previously mea-sured peak power. Server level power caps can be used to ensurethat the provisioned capacity is never exceeded and protect thecircuits and power distribution equipment.

Server level caps do not eliminate waste completely. Setting thecap at each server to its observed peak requires provisioning thedata center for the sum of the peaks, results in wasted capacitysince not all servers operate at the peak simultaneously. Instead,it is more efficient to provision for the peak of the sum of serverpower consumptions, or equivalently, the estimated peak powerusage of the entire data center. The estimate is based on previouslymeasured data and may sometimes be exceeded. Thus a cap mustbe enforced at the data center level. Here, the server level caps willchange dynamically with workloads. For instance, a server consum-ing a large amount of power need not be capped when some otherserver has left its power unused. However the former server mayhave to be capped when the other server starts using its fair share.Coordinated power capping systems [6–10] can be used for this.

Additionally, even the observed peak is only reached rarely. Toavoid provisioning for capacity that will be left unused most of thetime, data centers may provision for the 99th percentile of the peakpower. Capping would be required for 1% of the time, which may bean acceptable hit on performance in relation to cost savings. If thedifference in magnitude of power consumed at the peak and 99thpercentile is high, the savings can be significant. To quantify thesesavings, we present power consumption data from a section com-prising of several thousand servers in one of Microsoft’s commercialdata centers that host online applications serving millions of users,including indexing and email workloads. The solid line in Fig. 1shows the distribution of power usage, normalized with respect tothe peak consumption. If the 99th percentile of the observed peakis provisioned for, the savings in power capacity can be over 10%of the data center peak. Capacity reduction directly maps to cost

reductions.

Trends in server technology indicate that the margin for sav-ings will increase further. Power characteristics of newer serversaccentuate the difference between the peak and typical power

A.A. Bhattacharya et al. / Sustainable Computing:

Fig. 1. Cumulative distribution function (CDF) of power consumption for a clusteroco

(oao5lotia

2

tputaeiphc

bpftcfabc

2

sprscldp

f several thousand servers in one of Microsoft’s commercial data centers. Futureapacity reduction refers to the power consumed by the same workload if hostedn emerging technology based servers.

power consumed by a server under average load) usage becausef their lower idle power consumption. Power measurement for andvanced development server at different CPU utilizations showsnly 35% of peak consumption at idle, much lower than the over0% measured in current generation servers. Using processor uti-

izations from the real world servers, we project the power usagef the same workloads on the future generation servers assuminghat power scales with processor utilization [14] (the dashed curven Fig. 1). The present day data and technology trends both indicate

significant margin for savings.

.2.2. UPS chargingLarge data centers use battery backups, also referred to as Unin-

errupted Power Supplies (UPSs). UPSs provide a few minutes ofower during which time the diesel generators may be poweredp. After power is restored, the UPS consumes power to re-chargehe batteries. This implies that the power capacity provisioned for

data center should not only provide for the servers and coolingquipment but also include an additional margin for battery charg-ng. This additional capacity is almost always left unused sinceower failures are relatively rare. Even when power failures doappen, they may not occur at the time when data center poweronsumption is at its peak.

The capacity wasted due to reservation for battery charging cane avoided if the batteries are charged from the allocated serverower capacity itself. Should the servers happen to be using theirull capacity at recharging time, power capping is needed to reducehe server power consumption by a small amount and free upapacity for recharging batteries at a reasonable rate. Since powerailures are rare, the performance impact of this capping is accept-ble for many applications. Any data center that uses a batteryackup can use power capping to reduce the provisioned powerapacity.

.2.3. Total capital expensesMany power management methods are available to reduce

erver power consumption by turning servers off or using lowower modes when unused. Using less energy however does noteduce the cost of the power infrastructure or the servers them-elves. The amortized cost of the servers and power infrastructure

an be minimized if the servers are kept fully utilized [15]. Work-oad consolidation can help achieve this. Suppose a data center isesigned for a given high priority application and both servers andower are provisioned for the peak usage of that application. The

Informatics and Systems 3 (2013) 183– 193 185

peak workload is served only for a fraction of the day and capacityis left unused at other times. During those times, the infrastructurecan be used to host low priority applications. In this case cappingis required on power, as well as other computational resources,at all times to ensure that the low priority application is cappedto use only the resources left unused by the high priority applica-tions and up to a level that does not cause performance interferencewith the high priority tasks. Since power is capped by throttling thecomputational resources themselves, the implementation may notrequire an additional control knob for power. However, settings onthe throttling knobs should ensure that all resource limits and thepower limit are satisfied. The end result is that in situations wherelow priority workloads are available, power capping can be usedin conjunction with resource throttling to lower both power andserver capacity requirements.

2.2.4. Dynamic power availabilityThere are several situations where power availability changes

with time. For instance, if demand response pricing is offered, thedata center may wish to reduce its power consumption during peakprice hours. If the data center is powered wholly or partly throughrenewable energy sources such as solar or wind power, the avail-able power capacity will change over time. Power capacity mayfall due to brown-outs [16]. In this situation too, a power cappingmethod is required to track the available power capacity.

The above discussion shows that power capping can help savesignificant cost for data centers. However, existing power cappingmethods suffer from speed and stability limitations in certain prac-tical situations. In the next sections we quantitatively investigatethese issues and discuss techniques to enhance the existing meth-ods for providing a complete solution.

3. Speed: power capping latency

The actuation latency of power capping mechanisms is animportant consideration. Server level power capping mechanisms,typically implemented in server motherboard firmware, change theprocessor frequency using dynamic voltage and frequency scal-ing (DVFS) until the power consumption falls below the desiredlevel [4]. These local methods can operate very fast, typically cap-ping power within a few milliseconds. However, capping speedcan become an issue for coordinated power capping methodsthat dynamically adjust server caps across thousands or tens ofthousands of servers [9,8,10]. To understand this issue in depth,we first study the temporal characteristics of data center powervariations from the trace analyzed in Fig. 1. We then quantify therequired actuation latencies for a power capping mechanism, andcompare it to the state-of-the-art.

3.1. Data center power dynamics

Data center power consumption varies due to workload dynam-ics such as changes in the volume of requests served, resourceintensive activities such as data backup or index updates initiatedby the application, background OS tasks such as a disk scrubs orvirus scans, or other issues such as simultaneous server restarts.We study the data center power trace previously shown in Fig. 1 toquantify the rate of change of power.

Since capping is performed near peak power levels, only powerincreases that occur near peak usage matter for capping; powerchanges that are well below the peak, however fast, are not a con-cern. So we consider power increases that happen when power

consumption is greater than the 95th percentile of the peak. Wemeasure the rate of power increase, or slope, as the increase in nor-malized power consumption (if over the 95th percentile) during a10 s window.

186 A.A. Bhattacharya et al. / Sustainable Computing: Informatics and Systems 3 (2013) 183– 193

Fig. 2. Cumulative distribution function (CDF) of power slope [increase in powerconsumption of the cluster over a 10-s window]. The slope is normalized to thep

pistspc

3

fuw4cqtatccscaukcsatvtAotitc

TN

eak power consumed by the cluster during the period of the study.

Fig. 2 shows the CDF of the slope, normalized to the peakower consumption of the cluster. For most 10 s windows, power

ncreases are moderate (less than 2% of the peak cluster power con-umption). However, there exists power increases as high as 7% ofhe peak consumption over a 10 s window. To ensure protection andafety of electrical circuits during such extreme power surges, theower capping mechanism must be agile enough to reduce poweronsumption within a few seconds.

.2. Power control latency

This section experimentally investigates the limits on howast a power capping mechanism can throttle multiple serverssing DVFS. The experiments were performed on three serversith different processors: Intel Xeon L5520 (frequency 2.27 GHz,

cores), Intel Xeon L5640 (frequency 2.27 GHz, dual socket, 12ores with hyper-threading), and an AMD Opteron 2373EE (fre-uency 2.10 GHz, 8 cores with hyper-threading). All servers underest were running Windows Server 2008 R2. Power was measuredt fine time granularity using an Agilent 34411A digital multime-er placed in series with the server. The multimeter recorded directurrent values at a frequency of 1000 Hz, and root mean square wasomputed over discrete 20 ms intervals where one interval corre-ponds to 1 cycle of the 50 Hz AC power. Since in a practical powerapping situation, the cap will likely be enforced when the serversre busy, in our experiment the servers were kept close to 100%tilization by running a multi-threaded synthetic workload. Thisept the server near its peak consumption level from where powerould be reduced using power capping APIs. To estimate the fastestpeed at which a data center power capping mechanism can oper-te, the latency to be considered is the total delay in determininghe desired total data center power level, dividing it up into indi-idual server power levels, sending the command to each server,he server executing the power setting command via the relevantPI, and the actual power change taking effect (Fig. 3). Since we arenly interested in the lower limit on latency, we ignore the compu-ational delays in computing the caps. A central power controllers assumed to avoid additional delays due to hierarchical architec-ures. In the following sections we investigate each of these latency

omponents.

able 1etwork latencies in a data center.

Sender and receiver placement No. of samples Avg (ms) Std. dev (ms)

Within same rack 21 0.331 0.098Within same aggregation switch 32 0.342 0.030Under different aggregation switches 61 0.329 0.032

Fig. 3. Timeline showing the smallest set of latency components for a coordinatedpower capping solution. Additional latency components may get added when thecap is enforced in a hierarchical manner such as in [8,10].

3.2.1. Network latency in a data centerTable 1 shows the network latency of sending a packet between

the controller (hosted within the data center network) and thepower capping service at a server, for varying network distances.This data was obtained using a Microsoft data center managementtool, PingMesh, that allows measuring ICMP ping latencies across adata center network. The data shows that the average packet delayon a network is less than a millisecond. This latency component ishence not likely to be a concern for coordinated capping.

3.2.2. System latencyOnce a DVFS setting from the controller reaches a server, it is

applied by calling the relevant APIs. In this experiment, the fre-quency was decreased from the maximum to minimum to obtainthe highest resolution power change for measurement of latency.Low level frequency APIs offered through powerprof.dll in theWindows OS were used to avoid as much of the software stackdelays as possible. The threads for applying and reading the fre-quency setting were set to higher priority so as to not be delayeddue to the server workload. The latency incurred for changing thefrequency ranged between 10 ms and 50 ms for multiple runs onthe different servers.

3.2.3. Power change delayAfter the processor frequency changes there is an additional

delay before the power drops to the new level at the wall outlet, dueto factors such as capacitance in the server and the power supplycircuits. This effect requires a fine time granularity power measure-ment. Fig. 4 shows a sample power reading plot measured using theAgilent multimeter. The latency was found to be between 100 msand 300 ms, for a frequency change from the maximum to the min-imum, across multiple measurements over the three servers. Theminimum latency was observed when the frequency was changedbetween two adjacent DVFS levels requiring a smaller change inpower. The smallest latency across all adjacent frequency levels was60 ms. These measurements are similar to the fast capping latencyof 125 ms reported in commercial product data-sheets [17].

3.2.4. Total delayA summary of the latency results is provided in Table 2 and

totals to approximately 110–350 ms. This implies that for a feed-back based controller, it takes approximately 110–350 ms for oneiteration of a control loop. Much of this delay is coming from the

power change at the server itself rather than the computationaloverhead or network delay of the coordinated capping algorithm.

Implications: An important implication of the above measure-ments is that a feedback controller using multiple iterations can

A.A. Bhattacharya et al. / Sustainable Computing:

Fig. 4. Typical latency between the hardware frequency change and power reduc-td

isdftlnlfth

l

AntFvtwuo

t[bdcitet(

TS

ion. The current readings are rms values over discrete 20 ms windows. The powerecrease latency in this experiment was approximately 200 ms.

ncur several seconds of delay. Many controllers use a hierarchy tocale to a large number of servers or to logically separate the powerivision among multiple applications in a data center [8,10]. Wheneedback loops operate at multiple levels in the hierarchy, controlheoretic stability conditions require that the lower layer controloop must converge before an upper layer loop can move on to theext iteration. Suppose the actuation latency is denoted as l (where

≈ 110–350 ms from Table 2) and the number of iterations requiredor convergence at the ith layer in the control hierarchy is ni, thenhe total latency of the capping mechanism using N layers in theierarchy becomes:

total = l ×N∏

i=1

ni

s an example, considering the two layer hierarchy (N = 2) with1 = 6 and n2 = 16 used in [8], and plugging in the measured l value inhe above equation, we would get a control latency of 10.56–33.6 s.or the three layer hierarchy used in [10] and similar number of con-ergence iterations required, the latency will be even higher. Whilehis latency is not a concern for adapting to the slow changes inorkload levels that only cause the power to change every few min-tes, these latencies are not acceptable for the fast power changesbserved in real data centers (Fig. 2).

Some of the power distribution components in the data cen-er can handle capacity overages for a few seconds or even minutes18,19]. However, when power is changing at a rapid rate, the feed-ack based controllers cannot meet their stability conditions. Theynamics of the system being controlled must be slower than theonvergence time of the controller. The requirement for stabil-ty implies that power should not change beyond measurementolerance within the 10.56 s or 33.6 s control period. That how-

ver is not true since the power can change by as much as 7% ofhe data center peak power within just 10 s, in real data centersFig. 2).

able 2ummary of actuation latencies for power capping.

Latency type Approx. latency

Network <1 msOS 10–50 msWall power change 100–300 ms

Total 110–350 ms

Informatics and Systems 3 (2013) 183– 193 187

3.3. Summary

The latency analysis above implies that feedback based con-trollers using multiple iterations are not fast enough to operatesafely under the data center power dynamics. The design implica-tion for power capping methods is that the system may not havetime to iteratively refine its power setting after observing a capacityviolation.

Observation 1. A safe power capping system should use a single stepactuation to apply a power cap setting, such as using DVFS, that willconservatively bring the power down to well below the allowed limit(say, the lowest DVFS setting).

The conservative setting is needed to avoid unsafe operation inthe presence of model errors. Once power has been quickly reducedto a safe limit, feedback based controllers can be employed to itera-tively and gradually increase power to the maximum allowed limitto operate at the best performance feasible within the availablepower capacity.

4. Stability: application performance with DVFS basedcapping

It is well known that for system stability, the incoming requestrate should be lower than the sustained service rate across the mul-tiple servers hosting a given application [20]. This requirementis often the basis of capacity planning, such as for determiningthe number of servers required. The service rate is experimen-tally measured for a variety of requests served by the hosted onlineapplication and the number of servers is chosen to match or exceedthe maximum expected request rate.1 As request rate increases,more servers are added to the deployment.2 Under normal condi-tions, the service rate exceeds the request rate. However, wheneverpower capping is performed, power consumption of some serverresource must be scaled down. Typically the processor power isscaled down using DVFS for practical reasons, though in principle,one could scale down the number of servers or some other resourceas well. Regardless of the mechanism used to reduce power, engag-ing it reduces the service rate. The incoming request rate may ormay not change when service rate is reduced. If the system is closed,where each user submits a new request only after the previousresponse is received, the request rate will fall to match the servicerate. Batch processing systems such as Map-Reduce, HPC work-loads, or workloads such as mail where a user issues a new requestonly after a previous request is completed can be closely approxi-mated as a closed system. However, if the system is open, where therequest rate is not directly affected by the service rate, a decreasein service rate due to power capping may not lead to a equivalentdecrease in the request rate. Most web based online applications,such as web search, can be approximated as open systems sincethe requests are coming from a large number of users and newrequests may come from new users who have not yet experiencedthe reduced service rate. Even users experiencing reduced servicerate may not stop submitting new requests. Delays may even leadto rapid abort and retry.

4.1. Open-loop systems

Capping is enforced primarily when the system is at high powerconsumption. This happens when serving close to the peak demand

1 Power management methods may be employed to turn off or re-allocate unusedservers when request rate is lower than the maximum rate that can be served.

2 The terms request rate, demand and workload have been used interchangeablyin the subsequent sections.

188 A.A. Bhattacharya et al. / Sustainable Computing: Informatics and Systems 3 (2013) 183– 193

Fol

tcQid

uWi4uUawwelbssltvd

c

ig. 5. Server throughput and latency under different incoming request rates in anpen-loop system. (a) Application throughput against different incoming requestoads. (b) Increase in application latency as response rate increases.

hat the system can support. Hence, the reduced service rate afterapping is very likely to be lower than the demand at that time.ueuing theory says that response time shoots up uncontrollably

n this situation in an open-loop application. We experimentallyemonstrate this in an open system.

Experiment: We use a web server hosting Wikipedia pagessing MediaWiki,3 serving a 14 GB copy of the English version ofikipedia, as an example of an open-loop system. The front end

s hosted on an Intel Nehalem Xeon X5550 based server (16 cores,8 GB RAM) running Apache/2.2.14 and PHP 5.3.5. The databaseses mysql 14.14 hosted on another similar server. Both servers runbuntu 10.04 as the Operating System. HTTP requests for Apachere generated by a workload generator using the libevent library,hich has support for generating thread-safe HTTP requests. Sevenorkload generators were used to ensure that the workload gen-

rators themselves were not a bottleneck. To avoid crippling diskatencies, and to ensure that all requests were served out of theack-end cache the workload generators repeatedly request theame wiki page. All the workload generator servers have theirystem time synchronized and log performance (throughput andatency) in separate files that are then aggregated. We ramp uphe incoming request rate from 0 to 200 requests/s, to study theariation of delivered application throughput and latency at two

ifferent processor frequencies – 2.6 GHz and 1.6 GHz.

Observations: From Fig. 5(a), we find that the maximum appli-ation throughput remains equal to the incoming request rate at

3 http://www.mediawiki.org/.

Fig. 6. Effect of DVFS based power capping on throughput and latency in the exper-imental Wikipedia server. (a) Application average response time. (b) Applicationthroughput.

approximately 130 and 90 requests for 2.6 GHz and 1.6 GHz respec-tively. For larger magnitudes of incoming request rate, applicationthroughput becomes unpredictable. We thus conclude that withinthe above-mentioned frequencies, the application remains stable.From Fig. 5(b), we find that the average latency within the stablerange was 0.2 s. Incoming request rates beyond the stable region,also led to the request latency increase orders of magnitude (touch-ing a maximum of 20 s). This experiment clearly shows that drivingan open system with a request rate higher than its service rateresults is degraded performance.

Experiment: Next, we operate the system at a throughput of120 requests/s, which is below the maximum supported servicerate at 2.6 GHz. We conduct two experiments – one in which wekeep the server frequency at 2.6 GHz, and one in which we powercap the server to 1.6 GHz using DVFS, an existing power cappingmechanism, to reduce power consumption.

Observations: Fig. 6 shows the impact on performance usingDVFS based power capping. The gray curve shows the normaloperation at 2.6 GHz. The black curve shows the operation whenthe server is operated at a lower frequency but the incomingrequest rate is not changed. Throughput falls since the compu-tational resource available is lowered. However, latency starts toincrease uncontrollably to much higher values than the initial 0.2 s,even though the input request rate is constant throughout (at120 requests/s).

Performance plummets by orders of magnitude in a relativelyshort time when operating at the lower frequency. This is expectedsince several undesirable effects start to manifest in this situation.First, any buffers in the system, such as in the network stack or the

A.A. Bhattacharya et al. / Sustainable Computing:

Fai

wuiwibltlmoowfTs

4

coals

lwesstoipe

ddctdc

be served at a lower power level using the lower frequency, thoughthe peak throughput that can be served is lower at the lower fre-quency. A difference of 20 W is apparent. This is because the lower

ig. 7. Variation of throughput and latency when different processor caps arepplied to StockTrader. The processor cap was reduced by 10% every 60 s, as isndicated by the dotted vertical lines.

eb server’s application queue for incoming requests will get filledp and they will unnecessarily add to the latency without yield-

ng any advantage on throughput [21]. Second, requests not servedill be re-attempted, increasing the total number of requests com-

ng into the system. Since some of the requests served will note fresh requests but re-attempted ones, the total request service

atency will increase. Even with a small reduction in service rate,he number of dropped requests will start piling up and the averageatency will continue to rise, leading to the plummeting perfor-

ance observed. Third, if semantically, each user activity consistsf multiple requests (such as a accessing a web page may consistf accessing multiple embedded image and resource URLs from theeb server), since some of the requests may have been dropped

rom each semantic activity, no user activity will have been served.his implies that a small reduction in power can actually render aystem unstable.

.2. Closed-loop systems

In a closed-loop system, a reduction in service rate leads to aommensurate reduction in request rate because a new request isnly issued when the previous request terminates. While averagepplication latency does rise due to slower service rates, a closed-oop application does not experience the same cascading failure ashown in Section 4.1.

Experiment: As an example of a closed system, we use an instal-ation of Stocktrader 4.04 which mimics an online stock-trading

ebsite similar to IBM Websphere Trade 6.1. The workload gen-rators were set to behave as if in a closed system, where theyubmit the subsequent request only after the current request iserved. We use processor capping (reducing the running time ofhe application) to reduce the service rate and power consumptionf Stocktrader. An initial load of 2000 requests/s is applied, follow-ng which processor caps were used to throttle the application. Therocessor cap is reduced from 100% to 10% in decrements of 10%very 60 s.

Observations: Even though the application throughputecreases on the application of processor caps, Fig. 7 the latencyoes not rise uncontrollably. A lower service rate from the appli-ation server forces a reduction in the request rate generated by

he workload generators, resulting in no queue buildup or packetrops as seen in Section 4.1. Thus, application stability is notompromised during power capping of closed-loop systems.

4 http://msdn.microsoft.com/en-us/netframework/bb499684.aspx.

Informatics and Systems 3 (2013) 183– 193 189

Observation 2. A stable power capping system when reducing theservice capacity of a server running an open-loop application throughDVFS, should be able to implement a commensurate reduction inincoming application demand (through admission control).

5. Stable power capping with admission control inopen-loop systems

Admission control can be used with power capping mechanismsto reclaim stable behavior.

5.1. Admission control and power

Power capping reduces the service rate, which can make a sys-tem unstable. To maintain stability, the input request rate shouldalso be reduced within a modest time window, and admission con-trol is one technique to achieve that. This would result in some usersreceiving a “request failed” message or a long wait before service,but the system will be able to serve the remaining workload withinacceptable performance bounds.

If admission control is applied, the amount of work performed,and correspondingly the amount of computational resource used,is reduced. This implicitly reduces the power consumption sincethe processor has more idle cycles that it can spend in lowerpower sleep states. Intuitively, this suggests that admission con-trol can be used as an alternative power capping mechanism.We experimentally verify that this intuition is correct. However,there are practical issues that prevent admission control fromdirectly replacing DVFS based or other existing power cappingmechanisms.

Experiment: Using the same experimental testbed as usedin Section 4, we measure the power reduction provided usingadmission control. We implemented admission control using theiptables utility and selectively filter out requests from some of theworkload generators (based on IP address) to reduce the incomingrequest rate to the Wikipedia server.5

As in Fig. 6, suppose the server is originally operating at120 requests/s (at processor frequency 2.6 GHz). Suppose thedesired power reduction can be achieved using DVFS by lower-ing the processor frequency to 1.6 GHz. The throughput sustainedat this lower frequency is measured to be 85 requests/s andthe reduction in power is 46 W. Keeping the input request rateat 120 requests/s, we enforce admission control to allow only85 requests/s to be presented to the server. Fig. 8(a) shows theimpact on power when admission control is applied at the timetick of 140 s (approx). As intuitively expected, admission controldoes reduce power and can be used as a power capping mech-anism. However, the reduction in power is only 26 W (insteadof 46 W that was achieved using DVFS for the same reduction inthroughput).

5.2. Practical issues with admission control

Power efficiency: To investigate the power difference further,we measure the power consumption at varying throughput levelsat two different DVFS frequency settings. Fig. 9 shows the powermeasurements. The key take-away is that the same throughput can

5 In practice, admission control may be implemented by the application or in theload balancers, among other options. Our purpose in this paper is only to study theeffect of admission control on power and performance, and the above implementa-tion suffices.

190 A.A. Bhattacharya et al. / Sustainable Computing: Informatics and Systems 3 (2013) 183– 193

(a) (b)

F and adb th timd

fptwfl

isismtbifhm

orqtari

Fai

ig. 8. Power and average latency variation when using only admission control,

efore and after admission control is applied. (b) Variation of latency and power wiiscussed in Section 5.3.

requency is more energy efficient. As is known from DVFS design,rocessor power increases with the cube of frequency and even theotal server power has been measured to increase super-linearlyith frequency [20]. Since the number of processor cycles available

or computation increases only linearly with frequency, this makesower frequencies more energy efficient at a given throughput.

The observation above indicates that while admission controls required for stability, DVFS is more efficient from a power per-pective. Hence a practical power capping system must use DVFSn combination with admission control to achieve stability withoutacrificing efficiency. Fig. 8(b) shows the effect on power when bothechanisms are applied simultaneously (around time tick 163 s in

he figure). The throughput achieved (not shown) is the same inoth Fig. 8(a) and (b) but the power is capped by a greater amount

n Fig. 8(b). As technology improves and idle power consumptionalls further, the above power difference may be reduced since theigher frequency state with more idle cycles will likely becomeore power efficient as well.Delay: Another practical issue that requires DVFS is the effect

f queuing delays. If the application has a large buffer for incomingequests, then a large number of requests will be served from thatueue. Admission control will reduce the incoming request rate buthe service rate in the servers may remain high while the queues

re being emptied, leading to a delay before the power is actuallyeduced. This is a concern when speed of the capping mechanisms important.

ig. 9. Power vs throughput in the stable region of the Wikipedia front end servert 2.6 GHz and 1.6 GHz. Note the extra reduction in power available by using DVFSn addition to admission control

mission control + DVFS. (a) Power reduction and average request service latencye when admission control is used along with DVFS. The small increase in latency is

Safety: Admission control reduces the workload offered to theserver but does not force the server power to be lowered. Whilepower is expected to fall with reduced workload, in some cases itmay not, such as when the server is running a background virus scanor operating system update. With DVFS all computations related tothe workload or background tasks will be throttled down simulta-neously to reduce power.

5.3. Application latency

Another metric worth comparing between Fig. 8(a) and (b) isthe application performance in terms of latency. While throughputreduction is the same and stability is ensured in both cases, thelatency shows a small increase when DVFS and admission controlare combined. Suppose servicing each request requires an averageof nr processor cycles. Then the latency component attributable tothe processor, denoted lcpu can be computed as lcpu = nr/fi where fiis the processor frequency in use at the i−th DVFS setting. WhenDVFS is used to reduce the frequency from a higher value f0 to alower value f1, clearly lcpu will rise. Other latency components suchas the network round trip delay, queuing delay, and the latencyof accessing the backend storage are not significantly affected byDVFS and the increase in lcpu shows up as a small increase in overallapplication layer latency.

5.4. Summary

From the above analysis, we conclude that using admissioncontrol alone leads to a smaller power reduction, higher possibleactuation delay, and the possibility of unforseen software eventswhich might cause a power spike.

Observation 3. Admission control, while necessary for applicationstability, should be used in conjunction with DVFS to increase its effec-tiveness as a power capping knob.

The design implication is that power capping techniques shouldcoordinate with admission control agents, such as load balancers,to maintain application stability.

6. Admission control in closed-loop systems

Admission control is also applicable in a closed application sce-

nario, when the latency increase due to processor based powercapping mechanisms (DVFS or processor capping) is not desirable.Unlike an open system, a closed system remains stable when theservice rate is reduced (as shown in Section 4.2). Under a frequency

A.A. Bhattacharya et al. / Sustainable Computing:

Fig. 10. Difference in throughput and latency when admission control and processorcapping is applied to reduce power consumption in a server running the closed-loopSaa

sscclwutc

4Sabcr

tpoitlosls

tockTrader application. (a) Change in application latency for processor capping anddmission control. (b) Change in application throughput for processor capping anddmission control.

caled regime, both latency and throughput degrade due to thelower clock speed and (if applicable) additional buffering delay. Inontrast to DVFS, employing admission control could idle systemomponents to achieve the same power reduction while keepingatency unaffected. However, from the discussion in Section 5.2

e have seen that admission control has a lower throughput pernit power than DVFS. Thus, admission control trades off addi-ional throughput loss for its latency gains for power capping inlosed-loop systems.

Experiment: We use the same experiment setup as in Section.2. We reduce the power consumption of the server runningtockTrader from 89 W to 85 W through processor capping anddmission control. The network admission control was achievedy reducing the number of workload generators. Thus, we do notapture the additional load on the StockTrader server due to activeejection of requests.

Observations: Fig. 10 illustrates the use of admission con-rol in this closed-loop system and compares it to the use of arocessor based power capping mechanism. The system is firstperated at a throughput near 2100 requests/s where the latencys just below 0.04 s. At this point if the processor is throttledo reduce power, throughput falls to near 1500 requests/s whileatency rises to above 0.1 s (shown by the red arrows). On the

ther hand if the same power reduction is achieved through admis-ion control, throughput falls to around 1200 Frequest/s but theatency improves slightly due to lower number of requests beingerve (shown by the black arrows). Thus, in a closed-loop system,

Informatics and Systems 3 (2013) 183– 193 191

admission control provides improvement in latency (due to loadreduction), but reduces throughput. The exact nature of the tradeoffdepends on the specific application.

7. Issues in implementing admission control in distributedapplications

As mentioned in Section 5, a power capping action needs tobe accompanied by admission control to maintain application sta-bility. Admission control for open-loop replicated applications canbe performed efficiently at the load balancer, since it is the com-mon point of entry for all incoming requests. In this section, weidentify the need for coordinating the powercap controller (whichwould implement power caps) and the load balancer (which wouldimplement admission control).

7.1. Typical enterprise cluster configuration

Enterprise open-loop services are replicated and run on multipleservers behind a load balancer. The load balancer’s job is to dis-tribute new incoming connections between the replicated servers,employing algorithms such as weighted/non-weighted round-robin, least-connection scheduling (scheduling a new request tothe server currently serving the least number of requests), andsource hashing scheduling (scheduling a new request using a hashof the source IP address). Load balancers dynamically update serverweights to reflect server load profiles based on some periodicreporting mechanism [22]. Servers with higher weight receive ahigher number of incoming connections and vice versa. Thus, awell-tuned load balancer may infer power capping actions dynam-ically through its reporting mechanism, initiating the necessaryadmission control. This would remove the need for any coordina-tion between the powercap controller, and the load balancer.

7.2. Issues with lack of powercap-controller and load balancercoordination

Frequency of reporting mechanisms: There is a tradeoffbetween the frequency with which a load balancer updates serverweights and the amount of network traffic caused by the generationof this information. For instance, if the load balancer balances loadacross 100 servers, a reporting mechanism which provides updatesevery second generates 100 new network packets, increasing thenetwork traffic and affecting the performance of the load balancer(especially under overload conditions when a power capping actionmay be necessary). A slower rate of update, e.g., once every 5 s,reduces the number of control packets, but fails to inform a loadbalancer of possible server frequency scaling for a longer period oftime, during which the application could experience high latencies(as described in Section 4.1).

Incompatibiliy between powercapping and load balancingfeedback loops: When the power capping controller and theadmission controller do not coordinate (such as in [10]), the inter-action between a slow software power capping control loop andthe load balancer’s admission control loop should be studied forincompatibilities. Instabilities could crop up if the settling time ofthe software power capping control loop is comparable to the settlingtime of the load balancer control loop. Consider the following sce-nario: a power cap controller reduces the frequency of a server toreduce power consumption. The load balancer infers the reductionin server frequency through its reporting mechanism and allo-

cates less load to it. The lower load would drive down the powerconsumption of the server further. This might lead the powercapcontroller to increase the server frequency, creating a cycle, wherethe above-mentioned set of actions will be repeated. Coordination

1 ting:

mm

7

arAcworTosarnl

8

tpvcsdpmompaqbtc

dsptfmritothap

9

caitpe

[

[[[

[

[

[

[

[

92 A.A. Bhattacharya et al. / Sustainable Compu

echanisms such as the one mentioned in [23] could be used toitigate this incompatibility.

.3. An example implementation of coordination

A possible way to achieve stable coordination is that when capping action is enforced, the powercap controller instantlyeduces the weight of the affected servers through load balancerPIs, such as those provided by LVS. Additionally, the powercapontroller stops the reporting mechanism from modifying servereights (to prevent oscillation). While this could be achieved in

ur setup through temporarily stopping the Feedbackd daemonunning on the power-capped server, this technique is not general.he server weight for a power-capped server could be determinedffline through benchmarking, or online by dynamically analyzingerver load and power profiles. Under this particular coordinatedpproach, the need for frequent updates from the load balancereporting mechanism would be reduced because the load balancero longer infers power capping actions, and the possibility of oscil-

ations noted in Section 7.2 would be eliminated.

. Related work

Server level power capping methods [4] have been developedo throttle processor frequency in response to hardware meteredower readings at millisecond granularity. Similar techniques forirtualized servers have been investigated in [24,25], and use pro-essor utilization capping in addition to frequency control. Sinceingle servers methods do not make efficient use of the overallata center capacity, coordinated power budgeting across multi-le servers has also been considered [6,9,7,8,10]. We build on theseethods to address additional challenges. The coordinated meth-

ds rely on multiple feedback control iterations that, as we show,ay not satisfy convergence conditions under rapid data center

ower dynamics. Stability concerns with open-loop workloads arelso not considered in these works. The control of processor fre-uency in open and closed-loop system was considered in [20]ut for energy efficiency rather than power capping, and hencehe stability issue that arises in capping was not relevant in thatontext.

Admission control in web servers has also been studied inepth. Admission control methods drop requests to prevent theerver from getting overloaded [26–28] and maintain acceptableerformance. Feedback control and queuing theoretic algorithmshat carefully trade off the number of dropped requests and per-ormance have also been studied [29,30]. Processor frequency

anagement to maximize energy efficiency for variable incomingequest rates along with admission control have been consideredn [31,32]. Techniques to implement admission control by preven-ing new TCP connections or selectively blocking requests basedn the HTTP headers were presented in [33]. However, the integra-ion of processor frequency management and admission controlas not been considered for power capping. We discuss the desir-ble characteristics from both techniques that are relevant for thisroblem.

. Conclusions and future work

The cost of provisioning power and cooling capacity for dataenters is a significant fraction of their expense, often exceeding

third of the total capital and operating expense. Power capping

s an effective means to reduce the capacity required and alsoo adapt to changes in available capacity when demand responsericing or renewable energy sources are used. We described whyxisting methods for power capping lack two desirable properties

[

Informatics and Systems 3 (2013) 183– 193

of speed and stability and showed where these properties canmake the existing power capping mechanisms infeasible to beapplied. We also presented an approach based on admission controlto ensure stable and efficient operation. While admission con-trol cannot replace existing methods due to multiple practicalissues, we showed how it can provide the desirable characteristicsin a capping system when used together with existing mecha-nisms.

This work illustrates the usefulness of admission control inpower capping, but several open challenges remain. These includethe design of specific algorithms that control the extent of admis-sion control applied, its implementation in an efficient manner withminimal modifications of deployed applications, and safe integra-tion of multiple control mechanisms. Future work also includesprototyping rapid power capping mechanisms that can quicklyreduce power in case of rapid dynamics and then use feedbackto iteratively refine the power settings for maximum performancewithin the safe operating region. We believe that the understand-ing of relevant issues developed in this work will enable furtherresearch towards addressing these challenges.

Acknowledgement

This work was supported by the National Science Foundationunder grants CPS-0932209 and CPS-0931843.

References

[1] S. Pelley, D. Meisner, P. Zandevakili, T.F. Wenisch, J. Underwood, Power routing:dynamic power provisioning in the data center, in: ASPLOS, 2010.

[2] J. Hamilton, Cost of power in large-scale data centers, Blog entry dated11/28/2008 at http://perspectives.mvdirona.com, also in Keynote, at ACM SIG-METRICS 2009.

[3] L.A. Barroso, U. Holzle, The Datacenter as a Computer: An Introduction to theDesign of Warehouse-Scale Machines, Morgan and Claypool Publishers, 2009.

[4] C. Lefurgy, X. Wang, M. Ware, Server-level power control, in: Proceedingsof the Fourth International Conference on Autonomic Computing, Wash-ington, DC, USA: IEEE Computer Society, 2007, p. 4, Available from:http://dl.acm.org/citation.cfm?id=1270385.1270763

[5] P. Ranganathan, P. Leech, D. Irwin, J. Chase, H. Packard, Ensemble-level powermanagement for dense blade servers, in: Proceedings of the International Sym-posium on Computer Architecture (ISCA), 2006, pp. 66–77.

[6] M.E. Femal, V.W. Freeh, Boosting data center performance through non-uniform power allocation, in: ICAC, 2005.

[7] X. Wang, Y. Wang, Coordinating power control and performance managementfor virtualized server clusters, IEEE Transactions on Parallel and DistributedSystems 99 (2010).

[8] X. Wang, M. Chen, C. Lefurgy, T.W. Keller, Ship: a scalable hierarchical powercontrol architecture for large-scale data centers, IEEE Transactions on Paralleland Distributed Systems 23 (1) (2012) 168–176.

[9] R. Raghavendra, P. Ranganathan, V. Talwar, Z. Wang, X. Zhu, No “power”struggles: coordinated multi-level power management for the data center, in:ASPLOS, 2008.

10] H. Lim, A. Kansal, J. Liu, Power budgeting for virtualized data centers, in: USENIXAnnual Technical Conference, 2011.

11] D. Bhandarkar, Watt Matters in Energy Efficiency, Server Design Summit, 2010.12] Duke utility bill tariff. http://www.duke-energy.com/pdfs/scscheduleopt.pdf13] S. Govindan, A. Sivasubramaniam, B. Urgaonkar, Benefits and limitations of

tapping into stored energy for datacenters, in: International Symposium ofComputer Architecture (ISCA), 2011.

14] X. Fan, W. dietrich Weber, L.A. Barroso, Power provisioning for a warehouse-sized computer, in: Proceedings of ISCA, 2007.

15] J. Hamilton, Energy proportional datacenter networks, 2010. http://perspectives.mvdirona.com/2010/08/01/EnergyProportionalDatacenterNetworks.aspx

16] A. Verma, P. De, V. Mann, T. Nayak, A. Purohit, G. Dasgupta, R. Kothari, Brown-map: enforcing power budget in shared data centers, in: Middleware(ODP),December, 2010.

17] HP, Power regulator for proliant servers. http://h20000.www2.hp.com/bc/docs/support/SupportManual/c00300430/c00300430.pdf

18] X. Fu, X. Wang, C. Lefurgy, How much power oversubscription is safe andallowed in data centers? in: The 8th International Conference on Autonomic

Computing, June, 2011.

19] D. Meisner, T.F. Wenisch, Peak power modeling for data center servers withswitched-mode power supplies, in: Proceedings of the 16th ACM/IEEE Inter-national Symposium on Low Power Electronics and Design, Ser. ISLPED ’10,2010.

ting:

[

[

[[

[

[

[

[

[

[

[

[

[

[

A.A. Bhattacharya et al. / Sustainable Compu

20] A. Gandhi, M. Harchol-Balter, R. Das, C. Lefurgy, Optimal power allocation inserver farms, in: SIGMETRICS, 2009.

21] J. Gettys, K. Nichols, Bufferbloat: dark buffers in the internet Commu-nications of the ACM 55 (January (1)) (2012) 57–65, Available from:http://doi.acm.org/10.1145/2063176.2063196.

22] Feedbackd. http://jk.ozlabs.org/projects/feedbackd23] J. Heo, D. Henriksson, X. Liu, T.F. Abdelzaher, Integrating adaptive components:

an emerging challenge in performance-adaptive systems and a server farmcase-study, in: RTSS, 2007, pp. 227–238.

24] K. Rajamani, H. Hanson, J. Rubio, S. Ghiasi, L.R. F III, Application-aware powermanagement, in: IISWC, 2006.

25] R. Nathuji, P. England, P. Sharma, A. Singh, Feedback driven qos-aware powerbudgeting for virtualized servers, in: FeBID, 2009.

26] L. Cherkasova, P. Phaal, Session-based admission control: a mechanism for peakload management of commercial web sites, IEEE Transactions on Computers51 (June (6)) (2002) 669–685.

27] J. Carlstrom, R. Rom, Application-aware admission control and scheduling inweb servers, in: IEEE INFOCOM, 2002, pp. 506–515.

28] M. Welsh, D. Culler, Adaptive overload control for busy internet servers, in:Proceedings of the 4th Conference on USENIX Symposium on Internet Tech-nologies and Systems, vol. 4, ser. USITS’03, 2003.

29] X. Liu, J. Heo, L. Sha, X. Zhu, Adaptive control of multi-tiered web applicationsusing queueing predictor, in: NOMS, 2006, pp. 106–114.

30] M. Kihl, A. Robertsson, B. Wittenmark, Analysis of admission control mecha-nisms using non-linear control theory, in: ISCC, 2003, pp. 1306–1311.

31] V. Sharma, A. Thomas, T.F. Abdelzaher, K. Skadron, Z. Lu, Power-aware qosmanagement in web servers, in: RTSS, 2003.

32] C. Poussot-Vassal, M. Tanelli, M. Lovera, A control-theoretic approach for thecombined management of quality-of-service and energy in service centers, in:Run-time Models for Self-managing Systems and Applications, ser. AutonomicSystems, Springer Basel, 2010, pp. 73–96.

33] T. Voigt, P. Gunningberg, Adaptive resource-based web server admission con-trol, in: ISCC, 2002, pp. 219–224.

Arka Bhattacharya is a 3rd year Ph.D. student in Com-puter Science at the University of California, Berkeleyadvised by Prof. David Culler. He is currently pursuingresearch in enabling greater digital control of physicalbuildings. In the past he has worked on power-capping,energy proportionality and resource management in datacenters. Arka received his B.Tech in Computer Sciencefrom the Indian Institute of Technology Kharagpur in 2010where among other awards he received the Institute Sil-ver Medal, the Honda Young Engineer and Scientist InIndia Award (2008) and the O.P. Jindal Engineering andManagement Scholarship (2007).

David Culler is a Professor and Chair of Computer Science,Chair of Electrical Engineering and Computer Sciences,

and Faculty Director of i4energy at the University ofCalifornia, Berkeley. Professor Culler received his B.A. fromU.C. Berkeley in 1980, and M.S. and Ph.D. from MIT in1985 and 1989. He has been on the faculty at Berkeleysince 1989, where he holds the Howard Friesen Chair. He

Informatics and Systems 3 (2013) 183– 193 193

is a member of the National Academy of Engineering, an ACM Fellow, an IEEEFellow and was selected for ACMs Sigmod Outstanding Achievement Award, Sci-entific American’s ’Top 50 Researchers’, and Technology Review’s ’10 Technologiesthat Will Change the World’. He received the NSF Presidential Young Investiga-tors award in 1990 and the NSF Presidential Faculty Fellowship in 1992. He wasthe Principal Investigator of the DARPA Network Embedded Systems Technologyproject that created the open platform for wireless sensor networks based on TinyOS,and was co-founder and CTO of Arch Rock Corporation and the founding Direc-tor of Intel Research, Berkeley. He has done seminal work on networks of small,embedded wireless devices, planetary-scale internet services, parallel computerarchitecture, parallel programming languages, and high performance communica-tion, and including TinyOS, PlanetLab, Networks of Workstations (NOW), and ActiveMessages. He has served on Technical Advisory Boards for several companies, includ-ing People Power, Inktomi, ExpertCity (now CITRIX on-line), and DoCoMo, USA. Heis currently focused on utilizing information technology to address the energy prob-lem and is co-PI on the NSF CyberPhysical Systems projects LoCal and ActionWebsand PI on Software Defined Buildings.

Aman Kansal is a Researcher at Microsoft Research, in theSensing and Energy Research Group. He received his Ph.D.in Electrical Engineering from University of California LosAngeles, where he was honored with the department’sOutstanding Ph.D. Award. His current research interestsinclude computational energy efficiency in data centersand mobile computing. His research prototypes in theseareas have been recognized through international designcontest awards and are actively used worldwide, with oneof his recent prototype tools for energy efficiency exceed-ing two-hundred-thousand downloads. He has publishedover 65 research papers at premier computer science con-ferences and journals, and shipped his research through

Microsoft products including Windows Phone, Bing Mobile, and Visual Studio PhoneSDK. His work has also been recognized with the Microsoft Gold Star award, givenfor exceptional contributions towards Microsoft’s success. Dr. Kansal has served onthe NSF Committee of Visitors to review the NSF research funding process in Com-puter and Network Systems, co-chaired the PhoneSense and ImageSense workshopsat ACM Sensys, and served on numerous TPCs and organization committees.

Sriram Govindan received his B.Tech in information tech-nology at the College of Engineering Guindy, Chennai in2005, and his Ph.D. in computer science and engineeringat the Pennsylvania State University in 2011. He is a hard-ware engineer at the Datacenter Compute Infrastructurelab at Microsoft. His research interests include operatingsystem, server virtualization, performance characteriza-tion and datacenter infrastructure optimization.

Sriram Sankar is a Datacenter Architect in Microsoft’s

Online Services Division. His research interests are inDatacenter Architecture and Design Strategy, Workloadcharacterization, Operations Research and Emerging Stor-age Technologies. Sriram has a Masters degree from theUniversity of Virginia.