Efficient Resource Management for Cloud Computing Environments Andrew J. Younge1, Gregor von Laszewski1, Lizhe Wang1, Sonia Lopez-Alarcon2, Warren Carithers2 1: Pervasive Technology Institute Indiana University 2719 E. 10th Street Bloomington, Indiana 47408 2: Rochester Institute of Technology 102 Lomb Memorial Drive Rochester, New York 14623

Outline Introduction Motivation Related Work Green Cloud Framework VM Scheduling & Management Minimal Virtual Machine Images Conclusion & Future Work

What is Cloud Computing? “Computing may someday be organized as a public utility just as the telephone system is a public utility... The computer utility could become the basis of a new and important industry.” John McCarthy, 1961 “Cloud computing is a large-scale distributed computing paradigm that is driven by economies of scale, in which a pool of abstracted, virtualized, dynamically scalable, managed computing power, storage, platforms, and services are delivered on demand to external customers over the Internet.” Ian Foster, 2008 McCarthy described the vision of Utility computing, which has only become possible in recent history.

Virtualization Virtual Machine (VM) is a software artifact that executes other software as if it was running on a physical resource directly. Typically uses a Hypervisor or VMM which abstracts the hardware from an Operating System

Cloud Computing Features of Clouds Scalable Enhanced Quality of Service (QoS) Specialized and Customized Cost Effective Simplified User Interface Bottom line: Added functionality and reliability with small penalties

Data Center Power Consumption Currently it is estimated that servers consume 0.5% of the world’s total electricity usage. Closer to 1.2% when data center systems are factored into the equation. Server energy demand doubles every 4-6 years. This results in large amounts of CO2 produced by burning fossil fuels. What if we could reduce the energy used with minimal performance impact?

Motivation for Green Data Centers Economic New data centers run on the Megawatt scale, requiring millions of dollars to operate. Recently institutions are looking for new ways to reduce costs, no more “blank checks.” Many facilities are are at their peak operating envelope, and cannot expand without a new power source. Environmental 70% of the U.S. energy sources are fossil fuels. 2.8 billion tons of CO2 emitted each year from U.S. power plants. Sustainable energy sources are not ready. Need to reduce energy dependence until a more sustainable energy source is deployed. Just in the US alone, data centers are realistically responsible for about 50 million tons of CO2 each year. The U.S. makes up ¼ of the totally energy footprint and, with current globalization trends, this ratio will continue to shrink.

Green Computing Performance/Watt is not following Moore’s law. Advanced scheduling schemas to reduce energy consumption. Power aware Thermal aware Data center designs to reduce Power Usage Effectiveness. Cooling systems Rack design Performance has increased much more than performance per watt over the past few decades. PUE represents a metric of efficiency improving as the quotient decreases towards 1. Little research in designing efficient Cloud data centers

Research Opportunities There are a number of areas to explore in order to conserve energy within a Cloud environment. Schedule VMs to conserve energy. Management of both VMs and underlying infrastructure. Minimize operating inefficiencies for non-essential tasks. Optimize data center design.

Framework Green Cloud Framework Virtual Machine Controls Scheduling Power Aware Thermal Aware Management VM Image Design Migration Dynamic Shutdown Data Center Design Server & Rack Design Air Cond. & Recirculation

VM scheduling on Multi-core Systems There is a nonlinear relationship between the number of processes used and power consumption We can schedule VMs to take advantage of this relationship in order to conserve power Mension core i7 system Describe in more details, illustrating how I’ve created it Power consumption curve on an Intel Core i7 920 Server (4 cores, 8 virtual cores with Hyperthreading) Scheduling

Power-aware Scheduling Schedule as many VMs at once on a multi-core node. Greedy scheduling algorithm Keep track of cores on a given node Match vm requirements with node capacity Algorithm serves mainly as a template for further development. As jobs finish and relinquish there resources, pe is incremented. --Looking into scheduling based on expected execution time of VMs (TBD) Scheduling

485 Watts vs. 552 Watts VM VM VM VM VM VM VM VM Node 1 @ 170W Assume we have 4 nodes with the same power consumption curve described in slide 10 My Algorithm 1 compared to a classic Round Robin scheduling algorithm (used in Eucalytus and OpenNebula) VM VM Node 1 @ 138W Node 2 @ 138W VM VM VM VM Node 3 @ 138W Node 4 @ 138W

VM Management Monitor Cloud usage and load. When load decreases: Live migrate VMs to more utilized nodes. Shutdown unused nodes. When load increases: Use WOL to start up waiting nodes. Schedule new VMs to new nodes. Management

VM VM VM VM 1 Node 1 Node 2 VM VM VM VM VM 2 Node 1 Node 2 VM VM VM VM 3 Node 1 Node 2 VM VM VM VM 4 Node 1 Node 2 (offline)

Minimizing VM Instances Virtual machines are desktop-based. Lots of unwanted packages. Unneeded services. Are multi-application oriented, not service oriented. Clouds are based off of a Service Oriented Architecture. Need a custom lightweight Linux VM for service oriented science. Need to keep VM image as small as possible to reduce network latency. Management

Cloud Linux Image Start with Ubuntu 9.04. Remove all packages not required for base image. No X11 No Window Manager Minimalistic server install Can load language support on demand (via package manager) Readahead profiling utility. Reorder boot sequence Pre-fetch boot files on disk Minimize CPU idle time due to I/O delay Optimize Linux kernel. Built for Xen DomU No 3d graphics, no sound, minimalistic kernel Build modules within kernel directly This is a complex task! VM Image Design

Energy Savings Reduced boot times from 38 seconds to just 8 seconds. 30 seconds @ 250Watts is 2.08wh or .002kwh. In a small Cloud where 100 images are created every hour. Saves .2kwh of operation @ 15.2c per kwh. At 15.2c per kwh this saves $262.65 every year. In a production Cloud where 1000 images are created every minute. Saves 120kwh less every hour. At 15.2c per kwh this saves over 1 million dollars every year. Image size from 4GB to 635MB. Reduces time to perform live-migration. Can do better. FIX! VM Image Design

Conclusion Cloud computing is an emerging topic in Distributed Systems. Need to conserve energy wherever possible! Green Cloud Framework: Power-aware scheduling of VMs. Advanced VM & infrastructure management. Specialized VM Image. Small energy savings result in a large impact. Combining a number of different methods together can have a larger impact then when implemented separately.

Future Work Combine concepts of both Power-aware and Thermal-aware scheduling to minimize both energy and temperature. Integrated server, rack, and cooling strategies. Further improve VM Image minimization. Designing the next generation of Cloud computing systems to be more efficient.

Appendix

Cloud Computing Distributed Systems encompasses a wide variety of technologies Grid computing spans most areas and is becoming more mature. Clouds are an emerging technology, providing many of the same features as Grids without many of the potential pitfalls. From “Cloud Computing and Grid Computing 360-Degree Compared”

Data Center Design Need new data center designs strategies to reduce cooling requirements. Pod-based clusters: Modular Semi-portable Closed-loop systems Quebec’s CLUMEQ Silo supercomputer. Sun originally designed the blackbox, or Modular data center. Google has built warehouses to hook up such pods en mass. Reports a PUE of 1.2 to 1.4, depending on design. RIT plans to build a new data center to hold pods. CLUMEQ’s Silo design works well with a naturally cold environment

Minimal VM Image Ubuntu Linux Easier to slim down a fully functional distro than to create one from scratch. Selected Ubuntu Linux. Jaunty 9.04. Minimal install profile compared to other major distros. Excellent package management software (aptitude). Great support. FIX THIS SLIDE! Vs. Minimal Ubuntu VM Image Design

VM Scheduling Implemented scheduler on OpenNebula system Replaced Round Robin scheduling system with Based on Algorithm Startup and Shutdown VM Management Easily added From “Opennebula: The open source virtual machine manager for cluster computing”

Performance Impact of VMs Slight impact of scheduling 8 VMs instead of 4, but overall performance is still greater 4cores*93score=372score vs 8cores*60score=480score. This results in 22.5% more processing power when using all 8 hyperthreading cores. Especially interesting because it shows Hyperthreading can lead to a significant boost in performance, yet most data centers disable this feature.

DVFS VM Scheduling Image 1: Shows that the difference in power consumption between 2 VMs and 8 VMs is minimal, therefore the power consumption savings is considerable. Cost 11.5% to use all 8 vcores vs 4 Image 2: Slight impact of scheduling 8 VMs instead of 4, but overall performance is still greater 4*93=372 vs 8*60=480 22.5% more processing power