ICEP
Imperial College Energy and Performance Colloquium

Imperial College London, 29 May – 1 June 2012

Queen's Tower, Imperial College London

Timetable

ICEP talks will be held in Lecture Theatre 144 in Huxley Building (building 13 on the campus map) on the South Kensington Campus, apart from the EFL/ICEP invited evening lecture, which will be held in room G34 in the Sir Alexander Fleming Building (building 33 on the campus map).

Names in bold are invited talks, the others accompany an accepted extended abstract. Non-invited talks each have a half hour slot — we would expect the talk to last around 20 minutes to leave ample time for questions and transition to the next speaker.

Tuesday 29th May

Morning session
08:30 – 09:20 Registration and Tea/coffee Huxley 217/218
09:20 – 09:30 Jeremy Bradley
Imperial College London
Welcome
09:30 – 10:00 James Snee
University of Cambridge
Energy performance and accounting in mobile operating systems
Abstract: One obvious method of accounting for an application or subsystem's energy usage involves simply subtracting the energy usage of the device when running the application from idle consumption. For example, recent work has calculated the incremental energy cost of sending messages of varying size over a Wifi network by subtracting average idle power from power consumed under the chosen workload. But evidence shows that these systems aren't as entirely predictable as one would hope when accounting for energy in this way. Despite a clear trend the results include a number of prominent outliers. Given that this trend supports our expectations for energy consumption it seems likely that the outliers shown have resulted from non-representative experimental runs. For example, the preemptive multi-tasking nature of the underlying Linux OS could mean that the outlier points include energy costs from other, unrelated, activity. This presents two interesting research questions. First, where is the energy actually being consumed? And second, what causes these outliers? In this paper we describe our initial work on instrumenting the Android operating system. The goal is to account for energy consumption at the OS level, and to automatically identify these non-representative runs through unobtrusive monitoring of the system under test.
Slides: pdf
10:00 – 11:00 Ranveer Chandra
Microsoft Research
Smart offload to increase the battery life of mobile devices (part 1)
Abstract: Mobile devices are severely battery constrained. While smartphone capability has increased manifold in the last fifteen years, the battery energy density has only doubled. We have been exploring several techniques to improve the battery lifetime and in this talk I will discuss one such technique – offloading. In particular, by offloading computation from the main processor to either a lower power core, the network interface or to the cloud, we can put portions of the mobile device to sleep, thereby saving significant amount of energy.
Slides: pdf
11:00 – 11:30 Tea/coffee Huxley 217/218
11:30 – 12:00 Ranveer Chandra
Microsoft Research
Smart offload to increase the battery life of mobile devices (part 2)
12:00 – 12:30 Daniel Kelly
Imperial College London
Disaggregating multi-state appliances from smart meter data
Abstract: Smart electricity meters record the aggregate consumption of an entire building. However, appliance-level information is more useful than aggregate data for a variety of purposes including energy management and load forecasting. Disaggregation aims to decompose an aggregate signal into appliance-by-appliance information. Existing disaggregation systems tend to perform well for single-state appliances like toasters but perform less well for multi-state appliances like dish washers and tumble driers. In this paper, we propose an expressive probabilistic graphical modelling framework with two main design aims: 1) to represent and disaggregate multi-state appliances and 2) to use as many features from the smart meter signal as possible to maximise disaggregation performance.
12:30 – 13:00 Neil Davies
Predictable Network Solutions
Relating delivered experience to effects of underlying resource sharing
Abstract: Making choices as to how resources are shared requires an approach that helps decision makers make good, rational choices. Such approaches need to help inform likely outcomes given a particular resource allocation strategy as well as the other way round. This talk outlines the approach that has been used by us to discuss, plan and deliver more optimal use of another ephemeral resource, data network capacity, and its relationship with delivered quality of experience. Key in the approach is the creation of the boundaries that can be related to the business sectors so that the relevant stakeholders can understand their role, their added value and hence their likely income sources. We see that any optimisation of power will have to address similar issues in order to have a substantive effect.
Slides: pdf
13:00 – 14:00 Lunch Huxley 217/218
Afternoon session
14:00 – 15:00 Isi Mitrani
Newcastle University
Energy saving and performance in service centers
Abstract: We consider the problem of managing a service center where it is desirable to keep power consumption low, while at the same time maintaining high performance levels. These conflicting objectives are addressed by designating one or more blocks of servers as `reserves', to be powered up and down when the demand increases above or falls below certain thresholds. The questions of how to choose the paremeters of the operating policy are answered by analyzing suitable queueing models. These may allow customers to defect, i.e. to leave the system when their waiting times are too large. Simple and easily implementable heuristics are proposed and numerical results are presented.
Slides: pdf
15:00 – 15:30 Rossella Macchi
Politecnico di Milano
An energy-aware methodology for live placement of multi-mode virtual machines in the data center
Abstract: Desktop virtualization is a promising technology to improve the IT system management and provide a unified environment to the users. However, the users profile is usually overestimated in terms of allocated resources to guarantee the performance in every workload conditions. This usually leads to an unefficient use of resources and a huge energy consumption of the data center. This work proposes an approach to reduce the data center energy consumption by dynamically placing and dynamically adapting the performance profiles of virtual machines (VMs). We adopt two profiles with different resource requirements which the user can dynamically select based on the actual set of running applications. Then, we propose a mathematical approach and a heuristic approximated solution to efficiently place and migrate the virtual machines, while reducing the number of the active servers and the live VMs migrations. The results obtained on a realistic dataset took from a production environment show that the proposed approach is able to reduce the energy consumption of the data centers by almost 30% in average by proper allocating the virtual machines on the available servers.
Slides: pdf
15:30 – 16:00 Tea/coffee Huxley 217/218
16:00 – 17:00 Adam Wierman
Caltech
Algorithmic challenges for greening data centers
Abstract: Given the significant energy consumption of data centers, improving their energy efficiency is an important social problem. However, energy efficiency is necessary but not sufficient for sustainability, which demands reduced usage of energy from fossil fuels. In this talk, I will describe some recent work highlighting the algorithmic challenges associated with "greening" data centers. I will focus on two applications -- (i) dynamic resizing within a data center and (ii) geographical load balancing across an Internet-scale system. In both contexts I will present our new algorithms, which provide significantly improved performance guarantees when compared with the "standard" approaches using Receding Horizon Control. Additionally, if time allows, I will briefly discuss the our recent progress toward the implementation and evaluation of these algorithms in industry data centers.
Slides: pdf
17:00 – 17:30 Matthew Forshaw
Newcastle University
A novel approach to energy efficient content distribution with BitTorrent
Abstract: The energy efficiency aspects of IT infrastructure and communications systems are facing increased scrutiny, and a broad range of compelling financial, social, political and legislative factors is emerging. In this paper, energy efficiency considerations are addressed in the context of BitTorrent. We provide mechanisms to facilitate energy efficiency, namely Energy Proportional Tracker Migration, Peer Connectivity Shaping and Elastic Capacity Provisioning. We propose an energy-efficient content distribution system employing these mechanisms to minimise energy consumption and reduce cost.
Slides: pdf
18:30 – 20:30 ICEP Colloquium Reception at 170QG 170 Queen's Gate (for full delegates and invited speakers)

Wednesday 30th May

Morning session
08:30 – 09:00 Registration and Tea/coffee Huxley 344
09:00 – 10:00 Erol Gelenbe
Imperial College London
Dynamic energy management with/for smart ICT
Abstract: ICT is becoming one of the main culprits of CO2 emissions, already on a par with air travel, and soon probably doing even more damage. At the same time ICT is offering cleaner substitutes to emissions in areas such as transport because of the potential for substituting on-line activities for physical activities, such as working at home rather than commuting to an office. Furthermore ICT offers the potential to manage energy more efficiently and for better matching supply and demand, and substituting renewable energy sources in the place of fossil fuels. This lecture will address some of these questions from the perspective of classical computer systems performance engineering and show how some of our well established methods can be used to understand the trade-offs and help improve the outcomes.
10:00 – 11:00 Yuan Chen
HP Labs
Data center performance and power management (part 1)
Abstract: Data centers are very expensive to operate due to the power and cooling requirements of IT equipment. Rising energy costs, regulatory requirements and social concerns over green house gas emissions amplify the importance of energy efficiency. However, energy efficiency is for naught if the data center cannot deliver IT services according to predefined performance goals, as performance violations result in lost business revenue. Thus, an important question in data center resource management is how to correctly provision resource, such that performance requirements are met while minimizing energy consumption. This talk discusses several approaches in both theory and practice to improve the overall efficiency of data center operations.
Slides: pdf
11:00 – 11:30 Tea/coffee Huxley 344
11:30 – 12:30 Yuan Chen
HP Labs
Data center performance and power management (part 2) Slides: pdf
12:30 – 13:00 Samyukta Sethuraman
Texas A&M University
Application routing and server speed scaling for energy efficiency in data centers
Abstract: It is well-documented and accepted that data centers consume a phenomenal amount of energy. A significant portion of that can be reduced by resource management strategies such as assigning applications to servers, speed scaling, powering servers on or off and intelligent routing. Our group's previous study indicated tremendous gains by applying the above strategies in a unified and correlated fashion. However, solving a single model considering all these aspects is computationally intensive and in the current study, we propose a decomposed framework in which decisions at the lower level are made based on the decisions at the higher level. In this framework, decisions need to be made at multiple granularities. We consider a scenario in which a homogeneous set of servers serve a set of applications which bring demands according to a time varying load pattern. In the higher level, decisions are made for a longer time period like a day. At this level, the applications are assigned to servers according to correlations between the applications' arrival pattern. The objective is to have either strongly positively or strongly negatively correlated applications being served by the same server so that the server can either be turned off completely at some time periods in the first case, or can operate at a near-constant low speed in the second case. At the lower level, the decisions are made for a shorter time frame, for example an hour. The aggregate arrival rates over this time frame are collected. Routing decisions i.e. what fraction of arrivals of an application is routed to which server, are made by solving a deterministic optimization problem to minimize the energy utilized by the servers subject to QoS constraints. The decisions on server speeds are also made in this stage. Finally, we provide a fluid flow model to analyze the performance of the system under the decisions made above. In this analysis the stochastic system is modeled in real time as a semi-markov process.
Slides: pdf
13:00 – 14:00 Lunch Huxley 344
Afternoon session
14:00 – 14:30 Maryam Elahi
University of Calgary
Optimization problems in speed scaling systems
Abstract: Dynamic speed scaling has received significant attention in the recent literature as an effective method for balancing the need for higher speeds and the associated cost of energy. In this work, we investigate the tradeoffs between energy consumption and quality of service measures, such as response time and fairness in speed scaling systems. We first demonstrate that in a single server system with coupled speed scaling, where speeds are decided by a monotonically increasing function in the queue length, no policy can outperform the Processor Sharing in terms of response time and be fair to all jobs. We then introduce the notion of decoupled speed scaling, wherein the speed scaling function is not sensitive to the scheduling decisions, but adjusts the service rate based on the incoming load. Our initial simulation results suggest that in a model where there are costs associated with unfairness, decoupled speed scaling can be advantageous over the coupled speed scaling systems. Finally, we propose an objective function that combines the costs associated with different quality of service requirements under the umbrella of a general cost function. We aim to study the innate interactions between the scheduler and speed scaler under this model.
Slides: pdf
14:30 – 15:30 Evgenia Smirni
College of William and Mary
Autonomic exploration of trade-offs between power and performance in disk drives
Abstract: Over-provisioning is a standard capacity planning practice that leads to disk drives that operate mostly under very low utilization (as low as single digit utilization) but that are consuming disproportional amounts of power. Methodologies that place the disk drive into a low power mode during idle times can assist in conserving power. This is a challenging problem because the performance of future jobs cannot be compromised, yet there is no knowledge of future disk arrivals. In this paper we explore the above problem by exploring ranges and trade-offs of possible power savings and performance within a set of enterprise storage traces. We demonstrate the difficulty of obtaining significant power savings even in traces where overall utilization is less than 5% and explore the feasibility of popular schemes such as workload shaping for power savings. We also propose an autonomic algorithm that suggests when and for how long a power savings mode should be activated given an acceptable performance degradation target that is user provided. The robustness of the algorithm is illustrated via extensive experimentation.
Slides: pdf
15:30 – 16:00 Tea/coffee Huxley 344
16:00 – 17:00 Adam Wierman
Caltech
Energy procurement in the presence of intermittent sources
Abstract: The increasing penetration of intermittent, unpredictable renewable energy sources, such as wind energy, poses significant challenges for the utility companies trying to incorporate renewable energy into their portfolios. As a result, there is considerable discussion about how electricity markets should be restructured in order to facilitate the integration of renewable energy into the grid. Suggestions include adding additional markets, moving markets, closer to real-time, etc. In this talk, I will discuss how the optimal energy procurement of a utility company changes as a result of an increasing penetration of intermittent renewable resources, and what impact this should have on electricity market structure.
Slides: pdf
EFL/ICEP evening lecture / reception
18:00 – 19:00+ Francine Bennett
Mastodon C
Carbon reductions in computing: a pragmatic approach
Abstract: Cloud computing is a powerful tool which enables massive-scale data analysis and machine learning, but also has a very large carbon footprint, especially since many public cloud providers are powered by coal-fired grids (annual data centre emissions are currently ~75 million tonnes and growing). This talk discusses some practical approaches to reducing the carbon footprint of computing in data-led businesses.

Speaker biography: Francine is CEO and co-founder of Mastodon C, a cleantech company focusing on reducing carbon emissions in cloud computing without impact on cost or performance. Formed in February 2012, its goal is to help reduce global carbon emissions by 1%. Previously Francine has worked as Operations Director at Ask.com and Senior Financial Analyst at Google. She graduated from the University of Oxford in 2002 with a first class BA in Maths & Philosophy.

More information
Slides: pdf

Thursday 31st May

Morning session
09:00 – 09:30 Tea/coffee Huxley 344
09:30 – 10:00 Richard Hayden
Imperial College London
Mean-field models for interacting battery-powered devices
Abstract: We show how the mean-field approach can be extended to massively-parallel systems of interacting Markovian fluid models which can be used to describe networks consisting of many nodes, each powered and dependent on a local battery source. The mean-field limits manifest themselves as solutions to partial functional differential equations which admit inexpensive numerical solution using standard techniques. Results are compared here with stochastic simulation and show excellent agreement for reasonable population sizes.
Slides: pdf
10:00 – 11:00 Miklos Telek
Technical University of Budapest
Markov fluid models for energy and performance analysis (part 1)
Abstract: Markov fluid model is a flexible modeling tool to describe the behavior of system with hybrid (discrete and continuous) state space. One of the continuous variable characterizing the system behavior is the energy level. This way this modeling approach allows a combined analysis of system performance considering the energy consumption. The talk is going to survey the background of Markov fluid models and introduces some application examples.
Slides: pdf
11:00 – 11:30 Tea/coffee Huxley 344
11:30 – 12:30 Miklos Telek
Technical University of Budapest
Markov fluid models for energy and performance analysis (part 2)
12:30 – 13:00 Gareth Jones
Imperial College London
Using fluid queues to model energy storage and distribution
Abstract: Energy storage and distribution is an increasingly important problem as renewable (unreliable) energy sources supply more of our power needs. In this extended abstract we consider a simple network of energy suppliers, consumers and buffers modelled as fluid queues. Analytical evaluation of networks of fluid queues is hard as even tandem networks do no have a product form solution. We improve an approximation algorithm previously introduced to work in this new setting and evaluate our results using the Java JINQS simulator.
Slides: pdf
13:00 – 14:00 Lunch Huxley 344
Afternoon session
14:00 – 15:00 Evgenia Smirni
College of William and Mary
How to tame burstiness and save power in multi-tiered systems
Abstract: Burstiness (i.e. sudden surges) in user demands in enterprise systems that operate under the multi-tiered paradigm is a common phenomenon that leads to overprovisioning: the system is configured with excess hardware to meet peak user demands, often resulting in excessive (and unnecessary) power costs. In this talk, we present Fastrack, a parameter-free algorithm for dynamic resource provisioning that uses simple statistics to promptly distill information about changes in workload burstiness. This information, coupled with the application's end-to-end response times and system bottleneck characteristics, guides resource allocation, which proves to be effective under a broad variety of application burstiness profiles and bottleneck scenarios. Extensive simulations illustrate Fastrack's robustness for consistently meeting predefined service level objectives while minimizing power usage.
Slides: pdf
15:00 – 16:00 Guillaume Pierre
Vrije Universiteit
Performance control for complex Web applications (part 1)
Abstract: Dynamic resource provisioning aims at maintaining the end-to-end response time of a web application within a pre-defined range (Service Level Objective, or SLO). Provisioning resources for applications composed of multiple services remains a challenge. When the SLO is violated, one must decide which service(s) should be re-provisioned for optimal effect. We propose to assign an SLO only to the front-end service. Other services are not given any particular response time objectives. Services are autonomously responsible for their own provisioning operations and collaboratively negotiate performance objectives with each other to decide the provisioning service(s). After presenting the resource provisioning techniques themselves, I will discuss their application and implementation in the context of the ConPaaS runtime environment for elastic Cloud applications.
Slides: pdf
16:00 – 16:30 Tea/coffee Huxley 344
16:30 – 17:30 Guillaume Pierre
Vrije Universiteit
Performance control for complex Web applications (part 2)
19:30 – 21:00 ICEP Colloquium Dinner at ognisko 55 Exhibition Road (for full delegates and invited speakers)

Friday 1st June

Morning session
09:00 – 09:30 Tea/coffee Huxley 344
09:30 – 10:00 Lucia Gallina
Ca' Foscari University of Venice
A probabilistic energy-aware model for mobile ad-hoc networks
Abstract: Connectivity and energy consumption are two key aspects in mobile ad-hoc networks (MANETs). My research concerns a probabilistic, energy-aware, broadcast calculus for the analysis of mobile ad-hoc networks. It is based on a previous non-deterministic model, the EBUM (Energy-aware Broadcast, Unicast and Multicast Communications) calculus, which captures essential characteristics of MANETs, including the ability of a MANET node to broadcast a message to anyother node within its physical transmission range, and to move in and out of the transmission range of other nodes in the network. When dealing with MANETs, which use radio-frequencies to communicate, only broadcast transmissions are allowed, since channels can not be private. Moreover communications are not always addressed to the whole network, but a sender may be interested to reach a specific set of receivers (see, e.g., routing protocols, where each packet follows some specific path to reach the destination). This calculus, while considering the broadcast nature of communications, focuses the attention on the behaviour of the system with respect to the intended recipients of a message. The probabilistic version of the model extends the previous one by introducing probabilities modelling the mobility of nodes, that are not completely casual, but may depend on several aspects, such as the physical obstacles of the networks, or the nature of the devices ( e.g. the vehicular networks). Mobility of nodes is modelled by using Markov chains. The new calculus has been exploited to make performance analysis of mobile ad hoc networks, with a particular attention to the problem of energy conservation. The semantics is inspired by Segala's probabilistic automata driven by schedulers to resolve the nondeterministic choice among the probability distributions over target states. A probabilistic observational congruence has been provided, together with a bisimulation-based proof technique which allows one to verify whether two networks exhibit the same observable probabilistic behaviour (connectivity). Then an energy-aware preorder has been defined which can be exploited to compare the energy consumption of different, but behaviourally equivalent, networks. As an application, two well-known automatic repeat request (ARQ)-based error control protocols (stop-and-wait (SW) and go-back-N (GBN)) have been analysed and compared in terms of energy consumption.
Slides: pdf
10:00 – 11:00 Danilo Ardagna
Politecnico di Milano
Energy-aware autonomic resource allocation in multitier virtualized environments (part 1)
Abstract: In recent years, the energy consumption associated with Information Technology (IT) infrastructures has been steadily increasing. The reduction of energy usage is one of the primary goals of green computing, a new discipline and practice of using computing resources with a focus on the impact of IT on the environment. A significant amount of work has been done to achieve power reduction of hardware devices (e.g., in mobile systems to extend battery life). Nowadays, low power techniques and energy savings mechanisms are being introduced also in virtualized cloud data centres environments. In such systems, software is accessed as a service and computational capacity is provided on demand to many customers who share a pool of IT resources. Energy savings can be obtained by dynamically allocating computing resources among running applications and trading off application performance levels with energy consumption. As the customers access rates change significantly within a single business day, energy-aware resource allocation is a challenging problem and techniques able to control the system at multiple time scales are needed. In our research we have developed a framework addressing very large scale cloud data centres resource management. To validate its effectiveness, the proposed solution is compared to top-performing state-of-the-art techniques. The evaluation is based on simulation and on real experiments performed in a prototype environment. Synthetic as well as realistic workloads and a number of different scenarios of interest has been considered.
Slides: pdf
11:00 – 11:30 Tea/coffee Huxley 344
11:30 – 12:30 Danilo Ardagna
Politecnico di Milano
Energy-aware autonomic resource allocation in multitier virtualized environments (part 2)
12:30 – 13:00 Ioannis Dimitriou
Imperial College London
An unreliable vacation queueing model and its application on the DRX mechanism for power saving in 3GPP LTE
Abstract: The paper deals with an unreliable queue with modified vacations, timers, start-up, release period and its application to DRX mechanism for power saving mode in 3GPP LTE. The modified vacation scheme consists of two periods, a short DRX and a long DRX. Short DRX, has maximum number of N short cycles. If N consecutive short cycles end without arrival, the server enters the long DRX. We give expressions for the number of requests and its mean in steady state. Decomposition results are also presented. Energy metrics are obtained and used to provide useful numerical results.
Slides: pdf
13:00 – 13:30 Peter Thompson
Predictable Network Solutions
Creating change in large organisations – an experience report
Abstract: Increasing overall power "efficiency" is going to mean changed practices - changing practices means effecting change in organisations, having a large scale effect will require large organisations to change their practices in relative short order. Large organisations are not a single cohesive entity, it comes down to a complex interaction of people, practises, culture (both national and organisational) and, in no short measure, inertia. In this talk we will give examples of issues we have encountered when attempting to change perception and practices within large organisations, such as multi-nationals and government.
Slides: pdf
13:30 – 14:00 Lunch-to-go Huxley 344