Sky Computing on Future Grid and Grid 5000

Sky Computing on Future. Grid and Grid’ 5000 Pierre Riteau 1, Mauricio Tsugawa 2, Andrea Matsunaga 2, Jose Fortes 2, Tim Freeman 3, David La. Bissoniere 4, Kate Keahey 3, 4 1 Université de Rennes 1, IRISA/INRIA Rennes – Bretagne Atlantique 2 University of Florida 3 Argonne National Labs 4 University of Chicago Computation Institute Introduction n Architecture Sky computing is an emerging computing model where resources from multiple cloud providers are leveraged to create large scale distributed infrastructures. n n n This work uses resources across two experimental projects: Future. Grid and Grid’ 5000. This showcases not only the capabilities of the experimental platforms, but also their emerging collaboration. VM Image Propagation Mechanisms Our Sky Computing deployment makes use of: n n Xen to minimize platform (hardware and operating system stack) differences n Nimbus to offer VM provisioning and contextualization services (contextualization automatically assigns roles and configures VMs) n Vi. Ne, a virtual network based on an IP-overlay, to enable all-to-all communication between virtual machines spread across multiple clouds n Hadoop for parallel fault-tolerant execution and dynamic cluster extension Map. Reduce App (e. g. BLAST) The two platforms are used to create a Sky Computing environment. To validate our approach in a real-world scenario, we run a Map. Reduce version of a popular bioinformatics application (BLAST). However, any kind of distributed application can be run on these infrastructures. n To deploy virtual clusters, each VM requires an independent replica of a common VM image. Nimbus transfers a copy of the required VM image to each VM host (a step called propagation), using SCP from a single repository. This propagation scheme doesn’t scale with the number of VMs as it is limited by the repository disk or network bandwidth. To overcome this problem, we developed two new propagation mechanisms. The first one leverages the Tak. Tuk and Kastafior tools developed at INRIA to create a broadcast chain used to transfer image data. The second one relies on Copy-on-Write capabilities of the Xen hypervisor. Distributed Application (e. g. MPI BLAST) Hadoop Vi. Ne Nimbus Cloud A Nimbus Cloud B Cloud C Experimental Testbeds n n Future. Grid is an experimental testbed for grid and cloud research. It is distributed over 6 sites in the US and offers more than 5, 000 cores. Grid’ 5000 is an experimental testbed for research in large-scale parallel and distributed systems. It is distributed over 9 sites in France and offers more than 5, 500 cores. Scalability n n We deployed a Sky Computing infrastructure consisting of 1114 CPU cores (457 VMs) distributed over 3 sites in Future. Grid and 3 sites in Grid’ 5000 (OGF 29 demo, Chicago, IL, June 2010). Vi. Ne router San Diego Future. Grid’ 5000 University of Florida University of Chicago Grid’ 5000 firewall Rennes VMs The above graph compares instantiation times of virtual clusters using different propagation mechanisms. In the SCP and Tak. Tuk cases, the image is compressed and is 2. 2 GB in size (12 GB uncompressed). In the QCOW case, the 12 GB image is pre-propagated on all hypervisors. Propagation consists in creating a new Copy-On-Write volume and contextualizing the virtual cluster. Conclusion Grid’ 5000 Lille Queue Vi. Ne Router Sophia Sponsors and Acknowledgments This work is supported in part by the National Science Foundation under Grants No. OCI-0910812, IIP-0758596 and CNS-0821622 and in part by the MCS Division subprogram of the Office of Advanced Scientific Computing Research, Sci. DAC Program, Office of Science, U. S. Department of Energy, under Contract DE-AC 02 -06 CH 11357. The authors also acknowledge the support of the Bell. South Foundation. Any opinions, findings and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the National Science Foundation or Bell. South Foundation. Experiments were carried out using the Grid'5000 experimental testbed, being developed under the INRIA ALADDIN development action with support from CNRS, RENATER and several Universities as well as other funding bodies (see https: //www. grid 5000. fr). n The Sky Computing model allows the creation of large scale infrastructures using resources from multiple cloud providers. These infrastructures are able to run embarrassingly parallel computation with high performance. Our work shows how it is possible to federate multiple infrastructures and improve the speed of virtual cluster creation, using experimental testbeds in the US and in France as an example.