Energy Savings in the Data Center by Smart Controls on Load-Balancing - BBE2201 Spring 2018


 

 

Energy Savings in the Data Center by Smart Controls on Load-Balancing

Daniel Nissen nisse079
Senior Citizen Education Program
College of Continuing and Professional Studies
University of Minnesota
Major: Undeclared
BBE2201
2018 April 24
This paper reviews the literature on data center energy savings from redirecting services between data centers in different geographical locations.   The studies reviewed discuss the decision making process for allocating computational load to data centers, and the resulting energy savings.


Executive Summary

Introduction
Data centers use energy to compute, ventilate, light and for other activities.  They consume around 70 billion kWh per year or around 2% of the total energy use in the US.  The energy use stabilized over 2009-2014 as technology in the form of new chips, new storage devices, and virtual machines was introduced.  The large companies in this space have introduced renewable energy in several forms, but still consume considerable amounts of electricity from the grid.  The studies reviewed in this paper are focused on using more renewable energy by more intelligent assignment of computational tasks to data centers, and on making the prediction of how much energy from renewable sources will be available to do computational tasks in data centers.
Discussion
Many cloud service providers operate multiple data centers in separate geographical locations, and can move work between them at the speed of light.  If you choose carefully, you can allocate more work to the data centers with available renewable energy and less to data centers working with brown electricity (electricity from coal or natural gas, typically).  Data centers are usually overprovisioned with servers, so some of them can be asleep and not using much energy.  When needed they can be quickly awakened and the data needed for processing requests can be sent to them.  The studies reviewed here start by defining algorithms to decide how much renewable energy is likely to be available in each data center and then which tasks to assign to which data center.  Also, there are studies on how to predict wind and solar power availability that can be fed into the prediction models in the other studies.
Future predictions
Data center management will become more intelligent and complex as more non-dispatchable energy is integrated into the grid and installed by data center operators.  The cost savings can be considerable and visible.  Also, greenhouse gas reduction is becoming more popular and driving more decisions on which cloud provider to use.

Introduction and Background

The cloud computing revolution has led to data centers that use large amounts of energy, about 2% of all energy use in the US (Sverdlik, Y., 2016; Lawrence Berkeley National Laboratory, 2016).  The energy use stabilized from 2010 to 2014 (latest data) at about 70 billion kWh with improved efficiency balancing the large growth in computational capacity.  A small part of this energy actually is used by the chips to do computing.  Lighting uses even more energy in the data center.  Only about 14% of the capacity of most servers is actually used to do the computations (E. Source Companies, LLC, 2016).  This percentage is going up with better algorithms to do load-balancing and improved use of virtual machines and containers to move work around and on to fewer servers.
There are a lot of measures being implemented by data center operators to reduce energy use.  All energy used by the server computers is eventually converted to heat, so any reduction in server energy use is reflected in a reduction in cooling costs.  For every 1 kWh the servers need, about 0.6 kWh is needed to cool them (E. Source Companies, LLC, 2016). 
Google and Apple have both announced that they have Power Purchase Agreements or own their own renewable energy sources to power their needs for electricity (Google, Inc., n. d.; Sverdlik, Y, 2013).   But this does not mean they are always using renewable energy, only that the total capacity of renewables they have available matches or exceeds their requirements over time, maybe over a month or more.  Due to the non-dispatchable nature of renewable energy sources, they are using grid power at times that comes from other sources.  The research I am reviewing here leads to more use of renewable sources by moving work to locations more likely to have currently available renewable energy capacity.
The stabilization of the energy use by data centers in an era of increasing computational load is driven by improved efficiency.  A review of some commonly used techniques that are well known in the field may be worthy.  The use of energy by a server in a data center is relatively constant while it is powered up and running, whether it is actually performing user’s work or not.  Servers can be put to sleep when not needed, which significantly reduces the energy usage.  Disk drives can be spun down and up, and there are studies of how to position data on the drives to allow more time in the spun down state, significantly reducing energy use.  There is much literature on consolidating applications on fewer, busier servers, using virtual machines and containers to allow multiple application systems to coexist in a single hardware box.  Another way to reduce energy use may be to better control the temperature of the server components and allow the ambient temperatures around the server to rise. Another way to reduce energy use is to use alternative cooling methods, like locating the data center is a cold climate or using geothermal cooling.  All of these techniques reduce overall energy use in the data center and many are pretty standard in most data centers now (E. Source Companies LLC, 2016).  
The cost of energy for data centers varies widely due to different energy sources, and the availability of natural cooling due to a colder climate.  The marginal cost of energy may be significantly higher than the base load if that energy comes from the grid instead of local renewable sources (solar, geothermal or wind) (E. Source Companies LLC., 2016).
Availability of renewable energy varies significantly between locations due to amount of installed capacity and the weather, even over relatively short distances between data centers (Toosi, A. N., Qu, C., Assunção, M. D., & Buyya, R, 2017). 

Discussion of the Issue

One goal toward reducing our carbon footprint is to maximize the use of renewable energy and minimize the use of fossil fuel energy.  The other major driving force in looking at energy use is cost. 
The research I’m focusing on discusses various methods for moving workloads between geographical areas to increase the percentage of power used that comes from renewable sources.  Some of this research explores the use of forecasting methods to predict the availability of wind and/or solar power in each data center location over the next time period.  This predictability would allow the data center cluster to overprovision servers, and decide which servers to put to sleep and which to use for the computational load over the next time period.  Another approach monitors the availability of renewable power and periodically adjusts the load to the availability, in a reactive manner. The research on the models is either based on simulations or actual data captured from data centers.  None of these papers discusses an actual implementation in real data centers.
Predicting the availability of wind power is for the most part predicting the speed of the wind in a particular region. And predicting the availability of solar power is predicting the cloudiness in a region, as well as knowing sunrise and sundown times in the region. 
Using a reactive algorithm requires access to a source of renewable power availability (in watts) and knowledge of how many watts are needed to handle a particular load.  Khosravi, A., & Buyya, R. (2017) use a Gaussian Mixture Model (GMM) to predict short-term availability of renewable energy. This model uses current and previous availability figures for renewable energy to predict the future availability.  The researchers of this study did not implement a test system.  Instead they rely on historical data from the National Renewable Energy Laboratory (NREL) and workload demand from Amazon Web Services (AWS) to train their model.  The GMM model was able to predict up to 15 minutes with a 98% chance of getting within +/- 10% of the actual values.
Toosi, A. N., Qu, C., Assunção, M. D., & Buyya, R. (2017) describe another reactive model for load-balancing between data centers.  These researchers built a model for the allocation of workload to data centers with available renewable power, and used it in a real environment (Grid'5000 in France), as well as a simulation using traffic traces for English Wikipedia.  This model is an online algorithm, which means they do not know the future demand, or availability of renewable energy.  They introduce a global load balancer that routes requests to data centers where local load-balancers assign requests to individual servers.  Auto-scalers control whether to put servers to sleep or wake them up.  The Green Load Balancing (GreenLB) Policy is pretty straight forward.  Looking at available renewable energy and the price of brown energy at each data center, and thresholds of load the datacenters can each handle, they assign work to the lowest priced data center, treating renewable energy as free energy.  They always allow at least one server at each data center to be used, and so they may use some brown energy for that one server while other data centers still have available renewable energy.  Comparing this algorithm to the usual round-robin algorithm and another optimization technique, this study showed that this algorithm reduced the cost by 22% and 8% and brown energy by 17% and 8%, respectively. 
The major issue with these algorithms is how to optimize the prediction model.  Abedinia, O., Amjady, N., & Ghadimi, N. (2017) focus on predicting solar energy availability with a neural network approach.  They start by identifying the output of their algorithm, which is the predicted generation level of the photovoltaic cells.  The candidate inputs include solar radiation, temperature and photovoltaic (PV) generation each hour for the last 24 hours.  This is too many attributes for the neural network, so they use a two-stage feature selection method, focused on removing redundant and irrelevant inputs.  There is a complex set of 3 levels of neural network that process the inputs into the output.  The first neural network extracts a mapping function from the inputs, giving the resulting weights and output variable forecast as input to the second neural network. The second neural network does basically the same thing as the first, but with more precision, as its input is more precise.  The third neural network is similar.  The authors say you can continue this cascade but there is a limit on computational resources and a narrowing of the differences being detected, so they stopped at 3 neural networks.  These neural networks need to be trained, and the study used the Levenberg-Marquardt learning algorithm. The Shark Smell Optimization (SSO) is the unique algorithm introduced by these authors which is based on how sharks find their prey.  The algorithm tries to find increasingly dense odors, in this case increasingly consistent forecasts.  A major issue with neural networks is they trap at local minima, so this algorithm is designed to avoid that.  The end result of their neural network cascade is a significant reduction in the variance (normalized mean absolute percentage error and normalized root mean square error) over 9 other prediction methods.
Cheng, W. Y., Liu, Y., Bourgeois, A. J., Wu, Y., & Haupt, S. E. (2017) studied wind speed prediction and found that 0 to 3 hour forecasts mean absolute error could be improved by 30-40% by integrating anemometer data from the wind turbines into the prediction model.  Typically, wind turbines already include anemometers to allow them to adapt to conditions, like shutdown when the wind speed is too fast.  They started with a numerical weather prediction model, the Real-Time Four Dimensional Data Assimilation System (NCAR-ATEC RTFDDA or RTFDDA), and added the current wind speed and a calculated direction to it. This was a short (6-day) study and it needs expansion and more study. 
While the studies I have cited here are not definitive studies that show exactly how to build new systems to balance loads on data centers to use renewable power whenever possible, they combine to show progress toward that goal.  The overall meaning of these studies is that we can save significant energy and reduce costs by optimizing the usage of servers in data centers with available renewable energy on a 10 to 15 minute cycle. 

Future Predictions

We will continue to need to get smarter in order to manage the grid with renewable non-dispatchable energy and to reduce the cost of data center computation. The technology and processes to allow work to be moved between data centers and within data centers is well developed but can be continuously tuned and improved.  More data center operators are building renewable energy into their plans for future implementations.  The grid is also getting more renewable energy content all the time.  Future data center servers will continue to do more computation with less power.  Overall, of course, data centers will probably continue to use more energy, as we become more dependent on cloud services.  These optimization techniques are unlikely to improve faster than the growth of computational requirements.


 Bibliography

Abedinia, O., Amjady, N., & Ghadimi, N. (2017). Solar energy forecasting based on hybrid neural network and improved metaheuristic algorithm. Computational Intelligence,34(1), 241-260. doi:10.1111/coin.12145
Cheng, W. Y., Liu, Y., Bourgeois, A. J., Wu, Y., & Haupt, S. E. (2017). Short-term wind forecast of a data assimilation/weather forecasting system with wind turbine anemometer measurement assimilation. Renewable Energy, 107, 340-351. doi:10.1016/j.renene.2017.02.014
E. Source Companies LLC. (2016, January 21). Managing Energy Costs in Data Centers. Retrieved April 17, 2018, from https://ugi.bizenergyadvisor.com/data-centers
Google. Inc. (n. d.). Renewable energy – Data Centers – Google. Retrieved April 16, 2018, from https://www.google.com/about/datacenters/renewable/
Khosravi, A., & Buyya, R. (2017). Short-Term Prediction Model to Maximize Renewable Energy Usage in Cloud Data Centers. Sustainable Cloud and Energy Services,203-218. doi:10.1007/978-3-319-62238-5_8
Lawrence Berkeley National Laboratory. (2016, June). United States Data Center Energy Usage Report. Retrieved April 16, 2018, from https://eta.lbl.gov/publications/united-states-data-center-energy
Sverdlik, Y. (2013, March 21). Apple reaches 100% renewable energy across all data centers. Retrieved April 16, 2018, from http://www.datacenterdynamics.com/content-tracks/design-build/apple-reaches-100-renewable-energy-across-all-data-centers/74708.fullarticle
Sverdlik, Y. (2016, June 27). Here's How Much Energy All US Data Centers Consume. Retrieved April 16, 2018, from http://www.datacenterknowledge.com/archives/2016/06/27/heres-how-much-energy-all-us-data-centers-consume
Toosi, A. N., Qu, C., Assunção, M. D., & Buyya, R. (2017). Renewable-aware geographical load balancing of web applications for sustainable data centers. Journal of Network and Computer Applications,83, 155-168. doi:10.1016/j.jnca.2017.01.036


No comments: