Locality-Aware Cooperation for VM Scheduling in Distributed Clouds

The promotion of distributed Cloud Computing infrastructures as the next platform to deliver the Utility Computing paradigm, leads to new virtual machines (VMs) scheduling algorithms leveraging peer-to-peer approaches. Although these proposals considerably improve the scalability, leading to the management of hundreds of thousands of VMs over thousands of physical machines (PMs), they do not consider the network overhead introduced by multi-site infrastructures. This overhead can have a dramatic impact on the performance if there is no mechanism favoring intra-site v.s. inter-site manipulations. This paper introduces a new building block designed on top of a network with Vivaldi coordinates maximizing the locality criterion (i.e., efficient collaborations between PMs). We combined such a mechanism with DVMS, a large-scale virtual machine scheduler and showed its benefit by discussing several experiments performed on four distinct sites of the Grid' 5000 testbed. With our proposal and without changing the scheduling decision algorithm, the number of inter-site operations has been reduced by 72%. This result provides a glimpse of the promising future of using locality properties to improve the performance of massive distributed Cloud platforms.