Ankit Anand

Virtual Machine Placement Optimization Supporting Performance SLAs

Cloud computing model separates usage from ownership in terms of control on resource provisioning. Resources in the cloud are projected as a service and are realized using various service models like IaaS, PaaS and SaaS. In IaaS model, end users get to use a VM whose capacity they can specify but not the placement on a specific host or with which other VMs it can be co-hosted. Typically, the placement decisions happen based on the goals like minimizing the number of physical hosts to support a given set of VMs by satisfying each VMs capacity requirement. However, the role of the VMM usage to support I/O specific workloads inside a VM can make this capacity requirement incomplete. I/O workloads inside VMs require substantial VMM CPU cycles to support their performance. As a result, placement algorithms need to include the VMMs usage on a per VM basis. Secondly, cloud centers encounter situations wherein change in existing VMs capacity or launching of new VMs need to be considered during different placement intervals. Usually, this change is handled by migrating existing VMs to meet the goal of optimal placement. We argue that VM migration is not a trivial task and does include loss of performance during migration. We quantify this migration overhead based on the VMs workload type and include the same in placement problem. One of the goals of the placement algorithm is to reduce the VMs migration prospects, thereby reducing chances of performance loss during migration. This paper evaluates the existing ILP and First Fit Decreasing (FFD) algorithms to consider these constraints to arrive at placement decisions. We observe that ILP algorithm yields optimal results but needs long computing time even with parallel version. However, FFD heuristics are much faster and scalable algorithms that generate a sub-optimal solution, as compared to ILP, but in time-scales that are useful in real-time decision making. We also observe that including VM migration overheads in the placement algorithm results in a marginal increase in the number of physical hosts but a significant, of about 84 percent reduction in VM migration.

Resource usage monitoring for KVM based virtual machines

Realization of cloud computing has been possible due to availability of virtualization technologies on commodity platforms. Measuring resource usage on the virtualized servers is difficult because of the fact that the performance counters used for resource accounting are not virtualized. Hence, many of the prevalent virtualization technologies like Xen, VMware, KVM etc., use host specific CPU usage monitoring, which is coarse grained. In this paper, we present a performance monitoring tool for KVM based virtualized machines, which measures the CPU overhead incurred by the hypervisor on behalf of the virtual machine along-with the CPU usage of virtual machine itself. This fine-grained resource usage information, provided by the above tool, can be used for diverse situations like resource provisioning to support performance associated QoS requirements, identification of bottlenecks during VM placements, resource profiling of applications in cloud environments, etc. We demonstrate a use case of this tool by measuring the performance of web-servers hosted on a KVM based virtualized server.

Coarse-to-Fine Lifted MAP Inference in Computer Vision

There is a vast body of theoretical research on lifted inference in probabilistic graphical models (PGMs). However, few demonstrations exist where lifting is applied in conjunction with top of the line applied algorithms. We pursue the applicability of lifted inference for computer vision (CV), with the insight that a globally optimal (MAP) labeling will likely have the same label for two symmetric pixels. The success of our approach lies in efficiently handling a distinct unary potential on every node (pixel), typical of CV applications. This allows us to lift the large class of algorithms that model a CV problem via PGM inference. We propose a generic template for coarse-to-fine (C2F) inference in CV, which progressively refines an initial coarsely lifted PGM for varying quality-time trade-offs. We demonstrate the performance of C2F inference by developing lifted versions of two near state-of-the-art CV algorithms for stereo vision and interactive image segmentation. We find that, against flat algorithms, the lifted versions have a much superior anytime performance, without any loss in final solution quality.