1 / 51

ENERGY-EFFICIENT MANAGEMENT OF VIRTUAL MACHINES IN EUCALYPTUS

Pablo Graubner , Matthias Schmidt, Bernd Freisleben. ENERGY-EFFICIENT MANAGEMENT OF VIRTUAL MACHINES IN EUCALYPTUS. M Pazilijiang Feb 9, 2011. Outline. Background Problem Statement Saving Energy in Eucalyptus Implementation Experiment Results Conclusions.

armand
Download Presentation

ENERGY-EFFICIENT MANAGEMENT OF VIRTUAL MACHINES IN EUCALYPTUS

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Pablo Graubner, Matthias Schmidt, Bernd Freisleben ENERGY-EFFICIENT MANAGEMENT OF VIRTUAL MACHINES IN EUCALYPTUS M Pazilijiang Feb 9, 2011

  2. Outline • Background • Problem Statement • Saving Energy in Eucalyptus • Implementation • Experiment Results • Conclusions

  3. Background • Service demand increases(GeSI prediction) • Energy usage increases • Energy efficiency trend • Users turning to vendors • Vendors perusing energy efficiency Background Background System Model Problem Statement Saving energy in Eucalyptus Implementation Experimental Results Conclusions

  4. Background • Literature proposed energy efficiency in cloud computing, but • no full consideration of VM live migration in pre and post processing phases • Mostly theoretical or simulation-based • Effective energy efficiency should consider both • algorithm design • implementation Background Background System Model Problem Statement Saving energy in Eucalyptus Implementation Experimental Results Conclusions

  5. Background • Contribution • Research implemented in real cloud environment to improve the energy efficiency of infrastructure-as-a-service clouds. • Based on live VM migration to save energy • Takes into account pre and post processing phases of live migration • A multi-layered root file system used • Energy savings achieved up to 16% Background Background System Model Problem Statement Saving energy in Eucalyptus Implementation Experimental Results Conclusions

  6. Background • Eucalyptus • Transform your resources into Amazon EC2 compatible cloud computing system (http://aws.amazon.com/ec2) • Use existing EC2 client tools • Virtual Machine (VM) control • Network management • S3/EBS compatible storage • Open-source • Based on common open-source Linux/Java/Web service software • Can change/instrument code for focused experimentation Background Background System Model Problem Statement Saving energy in Eucalyptus Implementation Experimental Results Conclusions

  7. Background • Eucalyptus (continued) • Modular • Flexible enough to accommodate a variety of existing system configurations • Manageable • Highly configurable Background Background System Model Problem Statement Saving energy in Eucalyptus Implementation Experimental Results Conclusions

  8. Background (Source: www.ucgrid.org/summit/slides/dan-nurmi-eucalyptus.pptx) Background Background System Model Problem Statement Saving energy in Eucalyptus Implementation Experimental Results Conclusions

  9. Problem Statement • Practical problems that influence energy efficiency in cloud: • On-demand nature : unpredictable user behavior • Additional overhead: live migration and pre and post processing phases • Interference between VMs: recourse sharing in consolidated VMs Background Background Background Background System Model Problem Statement System Model Problem Statement Objectives Saving energy in Eucalyptus Single hop delay Experiment Case study Implementation Results End-to-end delay Experimental Results Discussion Results Conclusions Conclusions

  10. Problem Statement • Possible approaches for the practical problems • Optimizing the scheduling algorithm by considering/predicting different workloads of VMs or their execution time (might lead to overhead) • Optimizing the system in terms of energy efficiency (proposed in this paper, developing a solution that energy consumption and integrate into Eucalyptus) Background Background Background Background System Model Problem Statement System Model Problem Statement Objectives Saving energy in Eucalyptus Single hop delay Experiment Case study Implementation Results End-to-end delay Experimental Results Discussion Results Conclusions Conclusions

  11. Saving Energy in Eucalyptus • Power Management(PM) • PM requires an interface that is independent of hardware vendor specifications. ACPI (Advanced Configuration and Power Interface) [11] Background Background Background Background System Model Problem Statement System Model Problem Statement Objectives Saving energy in Eucalyptus Single hop delay Experiment Case study Implementation Results End-to-end delay Experimental Results Discussion Results Conclusions Conclusions

  12. Saving Energy in Eucalyptus • Storage Synchronization • Eucalyptus does not use NFS because of network traffic and overhead without live migration • DRBD (Distributed Replicated Block Devices) replicate data storage to different locations • DRBD module: stand-alone and synchronized • Stand-alone: disk access are passed to underlying disk driver. • Synchronized: disk writes are passed both to the underling disk driver and backup machine via TCP connection Background Background Background Background System Model Problem Statement System Model Problem Statement Objectives Saving energy in Eucalyptus Single hop delay Experiment Case study Implementation Results End-to-end delay Experimental Results Discussion Results Conclusions Conclusions

  13. Saving Energy in Eucalyptus • Storage Synchronization (Continued) • Eucalyptus uses MLRFS(multi-layered root file system) to synchronize the gigabit-sized disk images (Armbrust [4] stated data transfer bottlenecks as an obstacle for cloud computing) • MLRFS reduces the disk image transmission time • Disk image (downloaded from walrus) covers the read/write access to root file • Live migration (by Xen) synchronizes the virtual memory between source and destination host Background Background Background Background System Model Problem Statement System Model Problem Statement Objectives Saving energy in Eucalyptus Single hop delay Experiment Case study Implementation Results End-to-end delay Experimental Results Discussion Results Conclusions Conclusions

  14. Saving Energy in Eucalyptus • Eucalyptus Integration • Eucalyptus • Walrus • Cloud controller • Cluster controller • Node controller Background Background Background Background System Model Problem Statement System Model Problem Statement Objectives Saving energy in Eucalyptus Single hop delay Experiment Case study Implementation Results End-to-end delay Experimental Results Discussion Results Conclusions Conclusions

  15. Saving Energy in Eucalyptus • Eucalyptus Integration • Eucalyptus instance • Kernal/ramdisk • A root file system • Swap space • Network interface • Instance relocation • Relocation agent • Instance relocation algorithm Background Background Background Background System Model Problem Statement System Model Problem Statement Objectives Saving energy in Eucalyptus Single hop delay Experiment Case study Implementation Results End-to-end delay Experimental Results Discussion Results Conclusions Conclusions

  16. Saving Energy in Eucalyptus • Instance relocation • Relocation agent: • responsible for • Providing an interface enabling other components to initiate and abort instance relocations • Handling state transitions and events during the instance relocation process • Two major design goals for relocation agent • It should be reliable • It should be scalable Background Background Background Background System Model Problem Statement System Model Problem Statement Objectives Saving energy in Eucalyptus Single hop delay Experiment Case study Implementation Results End-to-end delay Experimental Results Discussion Results Conclusions Conclusions

  17. Saving Energy in Eucalyptus • Eucalyptus Integration • Instance relocation • Instance relocation algorithm • Follows a centralized server approach • Supervises the available set of servers and their recourses Background Background Background Background System Model Problem Statement System Model Problem Statement Objectives Saving energy in Eucalyptus Single hop delay Experiment Case study Implementation Results End-to-end delay Experimental Results Discussion Results Conclusions Conclusions

  18. Saving Energy in Eucalyptus • Instance Startup: The disk, kernel and ramdisk images are downloaded from Walrus. MLRFS support is added and the underlying DRBD devices are initialized, before the instance is started. VM relocation process Background Background Background Background System Model Problem Statement System Model Problem Statement Objectives Single hop delay Saving energy in Eucalyptus Experiment Case study Implementation Results End-to-end delay Experimental Results Discussion Results Conclusions Conclusions

  19. Saving Energy in Eucalyptus • Pre-Processing: On the destination server, the preparation of the instance relocation process is similar to an instance startup process: The images are downloaded from Walrus, MLRFS and DRBD devices are configured. VM relocation process Background Background Background Background System Model Problem Statement System Model Problem Statement Objectives Single hop delay Saving energy in Eucalyptus Experiment Case study Implementation Results End-to-end delay Experimental Results Discussion Results Conclusions Conclusions

  20. Saving Energy in Eucalyptus • Migration: The DRBD devices are synchronized, afterwards the live migration from source to destination server is performed. VM relocation process Background Background Background Background System Model Problem Statement System Model Problem Statement Objectives Single hop delay Saving energy in Eucalyptus Experiment Case study Implementation Results End-to-end delay Experimental Results Discussion Results Conclusions Conclusions

  21. Saving Energy in Eucalyptus • Post-Processing: On the source server, all remaining data is removed. On the destination server, the instance is activated and maintained by the destination Node Controller. VM relocation process Background Background Background Background System Model Problem Statement System Model Problem Statement Objectives Single hop delay Saving energy in Eucalyptus Experiment Case study Implementation Results End-to-end delay Experimental Results Discussion Results Conclusions Conclusions

  22. Implementation • Cluster Controller • Written in C programming language • Axis2/C Web Service operations to communicate with node controllers • Instance relocations are transparent to cloud controller • The instance relocation algorithm maps the servers and instance after the relocation is done • The relocation agent handles the state transition of relocations Background Background Background Background System Model Problem Statement System Model Problem Statement Objectives Saving energy in Eucalyptus Single hop delay Experiment Case study Implementation Results End-to-end delay Experimental Results Discussion Results Conclusions Conclusions

  23. Implementation • Node Controller • System replication configuration synchronizes the primary and secondary DRBD devices • The network bandwidth and TCP ports connecting DRBD are configurable • All MLRFS layers are overlaid using AUFS file systems Background Background Background Background System Model Problem Statement System Model Problem Statement Objectives Saving energy in Eucalyptus Single hop delay Experiment Case study Implementation Results End-to-end delay Experimental Results Discussion Results Conclusions Conclusions

  24. Implementation • Node Controller • Instance Startup • RunInstance operation • Pre-Processing • RelocationPreparation operation • Migration • RelocationMigration operation • Post-Processing • RelocationCommit operation Background Background Background Background System Model Problem Statement System Model Problem Statement Objectives Saving energy in Eucalyptus Single hop delay Experiment Case study Implementation Results End-to-end delay Experimental Results Discussion Results Conclusions Conclusions

  25. Implementation • Node Controller • Instance Startup • RunInstance operation is modified in order to mody the Xen config for multi-layer root file system support, swap ad ephermal disk creation for initializing DRBD • loopback devices are initialized, DRBD metadata is created. • Meta-disk consists of 128MB sparse image that is stored in meta-disk’s local user directory • Separate Perl script gen_libvirt_xml is used to generate VM config file Background Background Background Background System Model Problem Statement System Model Problem Statement Objectives Saving energy in Eucalyptus Single hop delay Experiment Case study Implementation Results End-to-end delay Experimental Results Discussion Results Conclusions Conclusions

  26. Implementation • Node Controller • Pre-Processing • RelocationPreparation operation is integrated into Node Controller web service and the instance state is transferred. • The root file system is downloded, Xen config is created, DRBD devices initialized. Background Background Background Background System Model Problem Statement System Model Problem Statement Objectives Saving energy in Eucalyptus Single hop delay Experiment Case study Implementation Results End-to-end delay Experimental Results Discussion Results Conclusions Conclusions

  27. Implementation • Node Controller • Migration • RelocationMigration operation signals to source to start • TCP ports are used to setup connections between primary and secondary DRBD and star synch. • virDomainMigrateTour() function to start migration Background Background Background Background System Model Problem Statement System Model Problem Statement Objectives Saving energy in Eucalyptus Single hop delay Experiment Case study Implementation Results End-to-end delay Experimental Results Discussion Results Conclusions Conclusions

  28. Implementation • Node Controller • Post-Processing • RelocationCommit operation signals source and target Node Controller • Source Node Controller the relocated instance removed • Target Node Controller the relocated instance added to cache • Recourses reallocated during the process freed • Remaining DRBD devices disconnected Background Background Background Background System Model Problem Statement System Model Problem Statement Objectives Saving energy in Eucalyptus Single hop delay Experiment Case study Implementation Results End-to-end delay Experimental Results Discussion Results Conclusions Conclusions

  29. Experimental Results • Experiment • Hardware : Intel CPU 3.0GHz, 4096MB RAM • Startup/Termination experiment: 2 servers • Others: 3 servers • EasyMeter for electricity measurement with UCM tool Background Background Background Background System Model Problem Statement System Model Problem Statement Objectives Saving energy in Eucalyptus Single hop delay Experiment Case study Implementation Results End-to-end delay Experimental Results Discussion Results Conclusions Conclusions

  30. Experimental Results • Experiment • Eucalyptus images Xen compatible • One virtual CPU, 348MB memory, 3GB disk space • Root file system size 1537MB • Ephemeral device size 1019MB • Additional 512MB swap space • DRBD synchronization rate of 50MByte/s Background Background Background Background System Model Problem Statement System Model Problem Statement Objectives Saving energy in Eucalyptus Single hop delay Experiment Case study Implementation Results End-to-end delay Experimental Results Discussion Results Conclusions Conclusions

  31. Experimental Results • Startup/Terminate Experiment • Intended to observe the power and energy consumption of an Eucalyptus cloud with randomly generated user request • Two measurement configuration: • Sleep interval of 1200 seconds • sleep interval of 2400 seconds Background Background Background Background System Model Problem Statement System Model Problem Statement Objectives Saving energy in Eucalyptus Single hop delay Experiment Case study Implementation Results End-to-end delay Experimental Results Discussion Results Conclusions Conclusions

  32. Experimental Results • Instance relocation algorithm does not lead to an increased energy consumption • One and two percent reduction is due to small differences in power consumption of the different servers. The energy efficient Eu coincidently relocated instance to a server with the smaller power consumption Energy Consumption of the Startup/Terminate experiment for two test configurations Background Background Background Background System Model Problem Statement System Model Problem Statement Objectives Single hop delay Saving energy in Eucalyptus Experiment Case study Implementation Results End-to-end delay Experimental Results Discussion Results Conclusions Conclusions

  33. Experimental Results • The difference is the number of instance started during the experiment(21 and 13) • First, the power consumption peak can be observed in correlation to the VM boot process during the start of an instance • Second, the number of instance running in the Config A > Config B. Config B has more suspended servers that lead to less power consumption Energy Consumption of the Startup/Terminate experiment for two test configurations Background Background Background Background System Model Problem Statement System Model Problem Statement Objectives Single hop delay Saving energy in Eucalyptus Experiment Case study Implementation Results End-to-end delay Experimental Results Discussion Results Conclusions Conclusions

  34. Experimental Results • MapReduce Experiment • Monitors energy consumption of an eucalyptus cloud during the execution of a distributed application • Apache Hadoop mapReduce implementation in combination with a Hadoop File System is used • The test application are installed and initiated by an external client automatically • Three servers are used. Server 1: hosts 2 instance(1 job tracker and 1 task tracker). Server 2: hosts 2 instance(1 task tracker and 1 idle instance). server 3: hosts 1 instance(1 task tracker) Background Background Background Background System Model Problem Statement System Model Problem Statement Objectives Saving energy in Eucalyptus Single hop delay Experiment Case study Implementation Results End-to-end delay Experimental Results Discussion Results Conclusions Conclusions

  35. Experimental Results • Wordcount experiment • Test configutation A: performed on text files with 110.4 MB total size 10 times in a row • Test configutation B: performed on text files with 110.4 MB total size 20 times in a row Background Background Background Background System Model Problem Statement System Model Problem Statement Objectives Saving energy in Eucalyptus Single hop delay Experiment Case study Implementation Results End-to-end delay Experimental Results Discussion Results Conclusions Conclusions

  36. Experimental Results • Config A: Energy efficient Ecu finishes 80s after server 3 is suspended, resulting reduced power consumption 8Wh, 10% of unmodified • Config B: saves 27Wh, 16% of unmodified Background Background Background Background System Model Problem Statement System Model Problem Statement Objectives Saving energy in Eucalyptus Single hop delay Experiment Case study Implementation Results End-to-end delay Experimental Results Discussion Results Conclusions Conclusions

  37. Experimental Results • Execition time: Config A: 5%, Config B: 7% • Difference is instance relocation process: synch and VM migration slows 30% Background Background Background Background System Model Problem Statement System Model Problem Statement Objectives Saving energy in Eucalyptus Single hop delay Experiment Case study Implementation Results End-to-end delay Experimental Results Discussion Results Conclusions Conclusions

  38. Experimental Results Shaded area is potential energy savings fir time t if the test was finished at t. There are 2 turning points: • Up until server 3 suspended • until the unmodified completed, the difference 35 Wh, later down to 27Wh Background Background Background Background System Model Problem Statement System Model Problem Statement Objectives Saving energy in Eucalyptus Single hop delay Experiment Case study Implementation Results End-to-end delay Experimental Results Discussion Results Conclusions Conclusions

  39. Experimental Results • Estimation of  • Test configutation A: PiEstimation is performed with 1,000 maps and 1,000,000 points • Test configutation B: PiEstimation is performed with 2,000 maps and 1,000,000 points Background Background Background Background System Model Problem Statement System Model Problem Statement Objectives Saving energy in Eucalyptus Single hop delay Experiment Case study Implementation Results End-to-end delay Experimental Results Discussion Results Conclusions Conclusions

  40. Experimental Results • Config A: Energy efficient Ecu finishes 60s after server 3 is suspended, resulting reduced power consumption 8Wh, 10% of unmodified • Config B: saves 27Wh, 16% of unmodified Background Background Background Background System Model Problem Statement System Model Problem Statement Single hop delay Objectives Saving energy in Eucalyptus Experiment Case study Implementation Results End-to-end delay Experimental Results Discussion Results Conclusions Conclusions

  41. Experimental Results • Execution time config A: 12%, config B: 6%, difference 130s Background Background Background Background System Model Problem Statement System Model Problem Statement Objectives Saving energy in Eucalyptus Single hop delay Experiment Case study Implementation Results End-to-end delay Experimental Results Discussion Results Conclusions Conclusions

  42. Experimental Results After server 3 suspended, the energy saving up to 32Wh Average Power consumption: Unmodified 286.5W Modified 250.3W,after server 3, it reduced to 235.3W Background Background Background Background System Model Problem Statement System Model Problem Statement Single hop delay Objectives Saving energy in Eucalyptus Experiment Case study Implementation Results End-to-end delay Experimental Results Discussion Results Conclusions Conclusions

  43. Experimental Results • KernCompile Experiment • A vanilla 2.6.33.1 kernel was compiled with the default configuration and four jobs compiling concurrently • The results are: • Energy consumed for an instance relocation is small • Confirmed the conclusion of the startup/terminate experiment • Instance relocation is not energy sensitive even though with DRBD synch and VM live migration Background Background Background Background System Model Problem Statement System Model Problem Statement Objectives Saving energy in Eucalyptus Single hop delay Experiment Case study Implementation Results End-to-end delay Experimental Results Discussion Results Conclusions Conclusions

  44. Experimental Results • Apache Benchmark Experiment • Aimed at emulating a web server application used by a cloud customer • The result is similar to KernCompile tests • Pattern of power consumption in the different phases is similar to kernCompile tests Background Background Background Background System Model Problem Statement System Model Problem Statement Objectives Saving energy in Eucalyptus Single hop delay Experiment Case study Implementation Results End-to-end delay Experimental Results Discussion Results Conclusions Conclusions

  45. Conclusion • An approach for improving the energy efficiency of infrastructure-as-a-service clouds. • It is based on VM live migration • Pre and post processing phases are considering during VM migration • DRBD, virtual file sytem • Short and long term test for evaluation • Up to 16% energy saving Background Background Background Background System Model Problem Statement System Model Problem Statement Objectives Saving energy in Eucalyptus Single hop delay Experiment Case study Implementation Results End-to-end delay Experimental Results Discussion Results Conclusions Conclusions

  46. References • Pablo Graubner et al, Energy-efficient Management of Virtual Machines in Eucalyptus, 2011 IEEE 4th International Conference on Cloud Computing. • Daniel Nurmi et al, Eucalyptus – Open- sources cloud computing system, www.ucgrid.org/summit/slides/dan-nurmi-eucalyptus.pptx

  47. Questions?

  48. Infrastructure-as-a-service Cloud infrastructure services, also known as "infrastructure as a service" (IaaS), deliver computer infrastructure – typically a platform virtualization environment – as a service, along with raw (block) storage and networking. Rather than purchasing servers, software, data-center space or network equipment, clients instead buy those resources as a fully outsourced service. http://en.wikipedia.org/wiki/Infrastructure_as_a_service#Infrastructure

  49. VM Live Migration • Migration of a virtual machine is simply moving the VM running on a physical machine (source node) to another physical machine (target node). • FAQ about live migration http://ppadala.net/blog/2010/06/understanding-live-migration-of-virtual-machines/ • Live Migration of Virtual Machineshttp://www.cl.cam.ac.uk/research/srg/netos/papers/2005-migration-nsdi-pre.pdf

  50. Contribution Quote “But in contrast to our approach, its implementation as part of a real-world cloud platform is only planned for the future. This is a fact that nearly all papers related to energy-efficient resource management in cloud computing have in common.” (Page 244)

More Related