Sky Computing on FutureGrid and Grid

1
Sky Computing on FutureGrid and Grid5000
Pierre Riteau1, Mauricio Tsugawa2, Andrea Matsunaga2, Jose Fortes2, Tim Freeman3, David LaBissoniere4, Kate Keahey3,4
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
Architecture
VM Image Propagation Mechanisms

• 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
  doesnt 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 TakTuk 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.

• Sky computing is an emerging computing model
  where resources from multiple cloud providers are
  leveraged to create large scale distributed
  infrastructures.

• Our Sky Computing deployment makes use of
• Xen to minimize platform (hardware and operating
  system stack) differences
• Nimbus to offer VM provisioning and
  contextualization services (contextualization
  automatically assigns roles and configures VMs)
• ViNe, a virtual network based on an IP-overlay,
  to enable all-to-all communication between
  virtual machines spread across multiple clouds
• Hadoop for parallel fault-tolerant execution and
  dynamic cluster extension

• This work uses resources across two experimental
  projects FutureGrid and Grid5000. This
  showcases not only the capabilities of the
  experimental platforms, but also their emerging
  collaboration.
• The two platforms are used to create a Sky
  Computing environment. To validate our approach
  in a real-world scenario, we run a MapReduce
  version of a popular bioinformatics application
  (BLAST). However, any kind of distributed
  application can be run on these infrastructures.

Distributed Application (e.g. MPI BLAST)
ViNe
Nimbus
Nimbus
Cloud A
Cloud B
Nimbus
Cloud C
Experimental Testbeds
Scalability

• FutureGrid 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.
• Grid5000 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.

• The above graph compares instantiation times of
  virtual clusters using different propagation
  mechanisms. In the SCP and TakTuk cases, the
  image is compressed and is 2.2GB in size (12 GB
  uncompressed). In the QCOW case, the 12GB image
  is pre-propagated on all hypervisors. Propagation
  consists in creating a new Copy-On-Write volume
  and contextualizing the virtual cluster.

• We deployed a Sky Computing infrastructure
  consisting of 1114 CPU cores (457 VMs)
  distributed over 3 sites in FutureGrid and 3
  sites in Grid5000 (OGF-29 demo, Chicago, IL,
  June 2010).

ViNe router
San Diego
Rennes
Grid5000 firewall
VMs
Conclusion
University of Florida
Lille

• 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.

Grid5000
FutureGrid
Queue ViNe Router
University of Chicago
Sophia
Grid5000
FutureGrid
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, SciDAC Program,
Office of Science, U.S. Department of Energy,
under Contract DE-AC02-06CH11357. The authors
also acknowledge the support of the BellSouth
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 BellSouth 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.grid5000.fr).