Efficient Resource Management for Cloud Computing Environments Andrew J. Younge, Gregor von Laszewski, Lizhe Wang, Sonia Lopez-Alarcon, Warren Carithers presented by Bryan Rosander Utility Computing • Long been a vision • Grid computing failed to really catch on • Technology advances as well as a viable business model have helped Cloud Computing catch on • Cloud Computing allows for fuller utilization of hardware • Energy consumption is turning into a major issue Is the Cloud Green? • 2005 o 0.5% of total world energy usage and 1.2% of U.S. energy usage come from data centers o World usage expected to quadruple by 2020, U.S. usage doubling every 5 years • More recent articles conflicting o Some suggest growth is slowing/has been slower (Reuters, Koomey) o Some suggest it is still increasing (Networkworld) Green Computing • In the past 15-20 years of supercomputers o performance has doubled > 3000 times o performance per watt has doubled 300 times o performance per square foot has doubled 65 times Scaling • Dynamic Voltage and Frequency Scaling (DVFS) o Intel SpeedStep o AMD PowerNow! • Started in laptops and mobile devices • Now used in servers Green Cloud Framework Green Cloud Framework (cont.) • Goal is to maximize performance per watt in a Cloud o VM Scheduling o VM Image Management o Data Center Design • Scheduling o Placement within cloud infrastructure o Energy use of server equipment, datacenter temperature important • Image Management o Small Size o Few unnecessary proce sses/services o Migration o Dynamic Shutdown • Data Center Design o More efficient A/C, power supplies o Hot and cold aisles o Utilizing external cooling Virtual Machine Scheduling • Thermal-Aware o Minimize overall temperature o Reduces energy used for cooling • Power-Aware o Minimize total power used by servers o Power to servers is the larger cost Virtual Machine Management • Can dynamically shutdown and start up machines as needed o Similar to Condor Glide-In (dynamically adds and removes machines from the resource pool) • Live migration can move virtual machines from lightly loaded to medium load servers o Can be used on machines idle during scheduling Virtual Machine Image • Operating systems are designed to run on diverse hardware o Not the case in the cloud o Normal for Linux to spend 15 seconds in modprobe o Reducing delay times, disabling modules can cut this down significantly • Graphical User Interfaces o Generally not necessary for cloud machines o Increase boot time o Increase size of image significantly • Boot order profile o Balance CPU utilization, I/O throughout entire boot o bootchart • Readahead Power Consumption Analysis • OpenNebula o open source distributed virtual machine manager o scheduler provides policies for virtual machine placement o Figure illustrates the CPU power savings (assuming CPU bound tasks) Virtual Machine Image Analysis • • • • • • Prototype Linux image created based on Ubuntu Linux 9.04 All unnecessary and desktop-oriented packages removed Image went from 4Gb to 636Mb Removed many daemons, processes, and libraries Utilized readahead to condense I/O into one burst Boot time went from 38 seconds to 8 seconds Conclusion • Power savings within the Cloud are an increasingly important area to focus on • Power-Aware scheduling can help increase utilization, synergizes well with dynamic shutdown and startup • Virtual Machine Image optimization can lead to gains on several fronts o Faster startup/shutdown increases effectiveness of dynamic startup/shutdown o Smaller images are easier to migrate, require less network traffic o Less wasted resources for the user Resources 1. Koomey, Jonathan - My new study of data center electricity use in 2010. http://www.koomey.com/post/8323374335 2. NetworkWorld - Report: Global data center energy use will rise nearly 20% next year. Chris Nerney. http://www.networkworld.com/newsletters/nsm/2011 /092611nsm2.html 3. Reuters - Data Center Power Use Drops as Green IT, Recession Take Effect. Iain Thompson. http://www.reuters.com/article/2011/08/02/idUS2 75708584920110802