Arunchandar Vasan

Worth their watts? - an empirical study of datacenter servers

The management of power consumption in datacenters has become an important problem. This needs a systematic evaluation of the as-is scenario to identify potential areas for improvement and quantify the impact of any strategy. We present a measurement study of a production datacenter from a joint perspective of power and performance at the individual server level. Our observations help correlate power consumption of production servers with their activity, and identify easily implementable improvements. We find that production servers are underutilized from an activity perspective; are overrated from a power perspective; execute temporally similar workloads over a granularity of weeks; do not idle efficiently; and have power consumptions that are well tracked by their CPU utilizations. Our measurements suggest the following steps for improvement: staggering periodic activities on servers; enabling deeper sleep states; and provisioning based on measurement.

Reducing data center power with server consolidation: Approximation and evaluation

With the growing costs of powering data centers, power management is gaining importance. Server consolidation in data centers, enabled by virtualization technologies, is becoming a popular option for organizations to reduce costs and improve manageability. While consolidation offers these benefits, it is important to ensure proper resource provisioning so that performance is not compromised. In addition to reducing the number of servers, there are other knobs — such as frequency/voltage scaling — that are being offered by recent hardware for finer granularity of power control. In this paper, we look at exploiting server consolidation and frequency/voltage control to reduce power consumption, while meeting certain provisioning guarantees. We formulate the problem as a variant of variable-sized bin packing. We show that the problem is NP-hard, and present an approximation algorithm for the same. The algorithm takes O(n2 log n) time for n workloads, and has a provable approximation ratio. Experimental evaluation shows that in practice our algorithm obtains solutions very close (< 6.5% difference) to optimal.

Thermal influence indices: Causality metrics for efficient exploration of data center cooling

Cooling is an important issue in data center design and operation. Accurate evaluation of a design or operational parameter choice for cooling is difficult as it requires several runs of computationally intensive Computational Fluid Dynamics (CFD) based models. Therefore there is need for an exploration method that does not incur enormous computation. In addition, the exploration should also provide insights that enable informed decision making. Given these twin goals of reduced computation and improved insights, we present a novel approach to data center cooling exploration. The key idea is to do a local search around the current design/operation of a data center to obtain better design/operation parameters subject to the desired constraints. To do this, all the microscopic information about airflow and temperature in data center available from a single run of CFD computation is converted into macroscopic metrics called influence indices. The influence indices, which characterize the causal relationship between heat sources and sinks, are used to refine the design/operation of the data center either manually or programmatically. New designs are evaluated with further CFD runs to compute new influence indices and the process is repeated to yield improved designs as per the computation budget available. We have carried out design exploration of a realistic data center using this methodology. Specifically, we considered maximization of the heat load in the data center subject to the constraints that: 1) servers are kept at appropriate temperatures and 2) overloading of CRACs is avoided. Our evaluation shows that the use of influence indices cuts down the exploration time by 80 % for a 1500 sq. ft. data center.

One meter to find them all: water network leak localization using a single flow meter

Leak localization is a major issue faced by water utilities worldwide. Leaks are ideally detected and localized by a network-wide metering infrastructure. However, in many utilities, in-network metering is minimally present at just the inlets of subnetworks called District Metering Area (DMA). We consider the problem of leak localization using data from a single flow meter placed at the inlet of a DMA. We use standard time-series based modeling to detect if a current meter reading is a leak or not, and if so, to estimate the excess flow. Conventional approaches use an a-priori fully calibrated hydraulic model to map the excess flow back to a set of candidate leak locations. However, obtaining an accurate hydraulic model is expensive and hence, beyond the reach of many water utilities. We present an alternate approach that exploits the network structure and static properties in a novel way. Specifically, we extend the use of centrality metrics to infrastructure domains and use these metrics to map from the excess leak flow to the candidate leak location(s). We evaluate our approach on benchmark water utility network topologies as well as on real data obtained from an European water utility. On benchmark topologies, the localization obtained by our method is comparable to that obtained from a complete hydraulic model. On a real-world network, we were able to localize two out of the three leaks whose data we had access to. Of these two cases, we find that the actual leak location was in the candidate set identified by our approach; further, the approach pruned as much as 78% of the DMA locations, indicating a high degree of localization.

Centralized Management of HVAC Energy in Large Multi-AHU Zones

HVAC control strategies that exploit temporal variations in zone occupancy have been well studied. However, at a given time, occupancy can also vary spatially within a single large zone with no internal wall partitions, that is served by multiple AHUs. We complement prior work by studying how spatial variations in a large zone can be leveraged to save energy and improve occupant comfort. Specifically, we propose a novel strategy for centralized reactive control of all the AHUs serving a large zone, MAZIC (Multi-AHU Zone Intelligent Control). To decide control outputs, we use a thermal model to capture the mixing of heat loads across different regions of the large zone served by different AHUs. We study MAZICs performance in terms of energy consumption and comfort using real-world occupancy data. When the spatial skew in occupancy is high, MAZIC reduces energy consumption by 11% over individual PID controllers running at each AHU, while maintaining similar comfort levels. Sensing temperature and occupancy at finer spatial resolution helps both MAZIC and PID controllers to save more energy when the occupancy is skewed. Finer spatial sensing does not add much value when the occupancy is not so skewed. We also find that augmenting MAZIC with a MPC (Model Predictive Control) approach yields insignificant improvement (< 3%) during normal occupancy. With ON-OFF occupancy patterns, MPC improves energy savings by up to 6% over reactive MAZIC.

Watts in the basket?: Energy Analysis of a Retail Chain

Electricity accounts for a significant part of a retail stores cost-to-serve. For a retail business spread across several stores, it is important to identify the correlations between cost, energy, operations, and location. To this end, we present a measurement-based analysis of energy and operations data gathered from 201 stores of a leading retail chain over a two year period. We employ statistical techniques and unsupervised learning to understand the inter-relationships across the various dimensions. Specifically, we find that: (i) The well-known Pareto cost-benefit principle (or the eighty-twenty effect) does not hold when considering the energy consumption as cost with customers served and store area covered as the benefits; (ii) After accounting for the time-of-day effects, sales counts do not affect energy consumption statistically, while ambient temperatures do so; (iii) Stores that exhibit a greater degree of energy proportionality have larger areas; and (iv) Opportunities for improvements exist in reducing the energy cost of operations. Many stores switch their operations on well ahead of their opening times. The average annual energy savings that could potentially be achieved across 201 stores if their operations are in tune with their opening time is roughly 8.2 GWh (2.5%). These savings can be achieved with just changes in operational procedures with zero capital investment.

Bugs in the Freezer: Detecting Faults in Supermarket Refrigeration Systems Using Energy Signals

Refrigeration is a major component of supermarket energy consumption. Ensuring faultless operation of refrigeration systems is essential from both economic and sustainability perspectives. Present day industry practises of monitoring refrigeration systems to detect operational anomalies have several drawbacks: (i) Over-dependence on human skills; (ii) Limited help in identifying the root-cause of the anomaly; and (iii) Presumption about high degree of instrumentation - which prevents their usage in supermarkets in developing economies. Existing approaches in literature to detect anomalies in refrigeration systems either are done in controlled laboratory settings or assume the availability of sensory information other than energy. In this paper, we present an approach to detect anomalous behavior in the operation of refrigeration systems by monitoring their energy signals alone. We test the performance of our approach using data collected from refrigeration systems across 25 stores of a real world supermarket chain. We find that using energy signal, we can not only detect anomalies but also narrow down the possible root-cause of the anomaly to a reduced set. Further, using energy signal along with data collected from other sensors (if available) allows us to reduce the false positive rate while identifying the root-cause of the anomaly.

Windy with a chance of profit: bid strategy and analysis for wind integration

Integration of wind power with the grid has become an important problem. For integration, a producer needs to bid in a time-ahead market to deliver an amount of energy at a future point in time. Because wind speed and price are both uncertain, a producer needs to place bids on the basis of expected wind power yield and price. To this end, improving the accuracy of the prediction of wind speed has received much attention. However, the trade-off between expected profit and the prediction errors over a multi-period setting has been less studied. We fill this gap by quantifying trade-offs between profits and prediction errors. First, we obtain, under idealized conditions on the price and the yield processes, an optimal bid strategy as a closed-form expression. Next, we evaluate the profit-vs-prediction trade-off using this idealized bidding strategy on synthetic datasets which satisfy all the idealistic assumptions. We also consider two baselines - a naive strategy and an oracle strategy that has perfect knowledge over a limited horizon. Finally, we relax our assumptions and evaluate all strategies under real-world datasets. We identify and work around limitations of the idealized bidding strategy when the underlying assumptions are violated. On synthetic datasets, with no buffering and a (relative) prediction error of 25\% , we find that our bidding approach performs significantly better than a naive approach and compares favourably (86\%) to an oracle with a look-ahead of two time-slots and infinite buffer. On real-world datasets, with buffer equivalent to 20\% of the maximum yield, our approach exceeds the naive approach by 25\%, while remaining within 62\% of a two-step look-ahead oracle that uses infinite buffering.

The Energy-Water Nexus in Campuses

Water is a critical index of an organizations sustainability. Since water reuse consumes energy, water management requires careful analysis of energy implications. To this end, we study the energy-water nexus in a multi-building campus with a water delivery network that spans multiple grades (such as potable, reclaimed sewage, etc). Using data collected over several months, we answer these questions: (i) What are the trade-offs between the external water footprint of a campus and its internal energy footprint of water? (ii) Are improvements in either footprint realizable in practice? (iii) Does reducing the consumption of one water grade have more impact on the energy consumption than other water grades? (iv) Does rainwater harvesting help reduce a facilitys energy footprint? We construct a multi-grade logical flow network with a per-link cost model for energy derived from the measured data. Under the constraint that demands are always met using the existing supplies, we optimize this flow-network for individually minimizing internal energy consumption of water and external water intake. Our study reveals the following: (i) minimizing external water footprint does not correspond to minimizing the internal energy footprint of water; (ii) demand reduction of different water grades impact the energy and water footprints differently; Contrary to intuition, reduction in second grade water demand yields highest reduction in water footprint while reduction in first grade water demand yields higher reduction in energy; (iii) Rainwater harvesting (RWH) can significantly reduce the energy footprint of a campus water network with sewage re-use. Our results show a potential for improving the operating condition of the campuss water network that can reduce the energy consumption by nearly 56 MWh (10.5%) and 99.6 MWh (18%) annually without and with RWH respectively.

On the Utility of Occupancy Sensing for Managing HVAC Energy in Large Zones

HVAC control strategies that exploit temporal variations in zone occupancy have been well studied. Occupancy can also vary spatially within a zone, especially during off-design operating conditions. We complement prior work by studying the usefulness of sensing occupancy information at different spatial resolutions in large zones served by multiple AHUs. As conventional PID controllers cannot utilize this information effectively, we propose a new control strategy and study the usefulness of sensing. We observe that utility of sensing occupancy at finer spatial resolutions is higher when the actual spatial heterogeneity in occupancy is higher.