Energy Efficiency and the Cloud

ENERGY EFFICIENCY AND THE CLOUD LECTURE 25 * 14-848 (CLOUD INFRASTRUCTURE) * FALL 2019

WHAT IS ENERGY? Energy, defined The capacity to do work Work, defined The transfer of energy from one body to another The product of a force and the amount of displacement.

WHY DO WE CARE BOUT ENERGY? Energy is the capacity to perform computation, store and access data, communicate, generate heat in the data center, move the heat in the data center somewhere else, etc. Energy is fueled by coal, natural gas, nuclear, hydroelectric, solar, wind, etc. The production of energy produces greenhouse gases , etc The transmission of energy is lossy, e.g. produces useless work, e.g. heating the environment The capacity to produce and transmit energy is limited and may be a limiting factor. Purchasing energy is an expense.

WHAT IS ENERGY EFFICIENCY? Classically, the ratio of useful work over total work Energizing a lightbulb may perform different types of work: Produce light Produce heat Produce a buzzing sound, e.g. florescent ballasts Only certain portions of the work are intended and useful E.g, producing light The others are wastage Producing heat and buzzing Energy efficiency is the amount of energy used for useful work / the amount of energy wasted on side-effects, e.g. Elight/(Eheat+ Esound)

WHY DO WE CARE ABOUT ENERGY EFFICIENCY? Cost Energy must be purchased Limitation There are limits to the amount of energy that can be bought and distributed Capacity for work may be limited by available energy Environmental impact Land and emissions from fossil fuels Waste storage, potential for release from nuclear energy Dead birds, lost land or water/navigability, landscape/horizon from wind turbines Pollution from production and retirement of solar Loss of nature site and fish habitat from hydroelectric Etc. Complexity Side-effects, such as heat production, often require additional systems and support to manage them, e.g. cooling.

HOW MUCH ENERGY DO DATA CENTERS USE? Various estimates from 1% - 1.5% of total energy consumption Varies by region and estimate Corresponding amount of carbon emissions Even larger if considered more broadly than just data center, e.g. telecom infrastructure Many regions are energy limited No more power to buy Limited production Limited distribution Industrialized region problem, not just developing regions

HOW MUCH ENERGY DO DATA CENTERS USE: BY COMPARISON Typical house in US uses 10,000 kw-h/year 500 m2(5,382 ft2) of data center 25,000 30,000 kw-h/day 10,000,000 kw-h/year Equivalent to about 1,000 US households Equivalent to about 2,000 European households Equivalent to about 8,000 Warm climate households

WHERE DOES THE ENERGY GO? Estimates vary widely Depends a lot on the data center age, configuration and management techniques For some sense Cooling ~15-50% (Varies by design, temperature, etc) Servers ~33% Networking gear~10% Storage 15% (Totally depends upon purpose of data center and amount of storage and SSD vs Disk) SSD uses about 25% as much energy as disk, but varies dramatically with load, e.g. read vs write, idle vs working idle and read are much, much cheaper than disk Loss to power conversion 5-10%

NATURAL TENSION HIGH AVAILABILITY VS ENERGY EFFICIENCY Energy efficiency favors less equipment operating at higher work density Less equipment consuming energy Availability favors more equipment operating at lower work density More redundancy Less at risk per failure Greater ability to handle compounding bursts

COMMON GOALS Improve efficiency E.g. Energy converted vs leaked by power supplies Leaked energy is, for example, often lost as heat and then requires cooling Reduce overhead Loss due to partitioning, e.g. multiple small power or cooling units instead of a larger one, etc. Reduce idle time Consuming energy and not getting any benefit isn t good for efficiency! Reduce redundancy In hardware In software processes

ENERGY SAVING TECHNIQUES Managing heat Manging hardware & software efficiency Workload prediction VM placement VM migration Workload consolidation Resource overcommitment Network elasticity, lack thereof Cloud management

MANAGING HEAT Idle servers produce heat Running and accomplishing nothing is a waste of energy That wasted energy is often heat and then requires cooling, which requires more energy Fans may not be efficient More ventilation than needed Moving more air than needed because warm air as input due to rack and center configuration Thermoelectric cooling is often more efficient for point loads Hard to get enough airflow in only a small area results in greater, wasted, airflow over larger area

MANAGING HARDWARE & SOFTWARE EFFICIENCY Oversized caches in systems with poor locality waste memory and thereby energy Appliances may impose less OS wastage than general purpose OS Unused services not running and consuming energy. GPUs tend to be more energy efficient than CPUs

WORKLOAD PREDICTION Physical machines consume energy whether doing useful work or not Idle servers have been reported to consume about 50% as much energy as at peak load (utilization) Switching systems in-and out of sleep state requires both time and energy Storing and restoring data, etc. Cost can exceed benefit for short sleep periods Prediction needed Don t sleep if needed again soon Turn on before needed to avoid latency Save energy cost of unhelpful sleep-wake cycle

VM PLACEMENT CLUSTER SELECTION Cost of energy varies by region and time Ability to predict energy costs needed for good placement Consideration can be given to sources providing power, e.g. coal vs. wind, solar, hydroelectric, etc.

VM PLACEMENT HOST SELECTION Want hosts as full as possible without going over Low density means more hosts than needed Remember high cost of idle More hosts means more energy, capital, operation expenses Density too high Missed service guarantees Possibly greater energy cost, e.g. disk swapping Viewed as bin packing problem NP hard, but good approximation strategies

VM MIGRATION Concentrate workloads by migrating sparsely distributed workloads to a denser configuration Encounters energy cost due to extra work for migration Only makes sense to migrate VMs that will last Predict older ones Others can be left to expire. Sleep unused hosts once workloads are concentrated.

OVERCOMMITMENT Typical numbers suggest About 1/3 of requested CPU resources are used About of requested memory is needed Consequences More resources powered up than actually in use. Cause Clients don t really know what they need, so they request more to be safe Bursty loads set the ceiling, but don t keep it there. Solution Overcommit, taking advantage of unlikelihood of stacking Migrate if it gets bads

NETWORK ELASTICITY Switches consume a greater percentage of their energy at idle than servers Up to 80% energy used at idle. Servers and storage are often throttled back Shut down or put to sleep But, network tends to be less elastic This means as utilization goes up due to concentration, the network accounts for more of the energy consumption Rule of thumb: The higher the utilization, the greater the relative network cost There isn t a natural reason for this, and network gear can be reconfigured and shut down like servers Network switches may be able to sleep interfaces, but commonly can t sleep backbone. But, may not be possible, depending upon demand, architecture, and ability to manage

NETWORK ELASTICITY Bursty loads are the enemy of network energy efficiency Designing networks for energy efficiency reduces headroom Reduced headroom is a problem for bursts Certain types of long-lived, predictable loads are more manageable Predictable times of day and days of week for utilization Examples: Streaming entertainment video, gaming, teleconferences, etc