Higher consolidation ratios in virtualized datacenters, SSD based storage arrays and converged IO fabric are mandating the need for more flexible and powerful models for physical network adapter bandwidth allocation. Existing solutions such as fair-queuing mechanisms, traffic-shapers and hierarchical allocation techniques are insufficient in terms of dealing with various use cases for a cloud service provider.

Current public cloud offerings sell capacity in the form of pre-defined virtual machine (VM) configurations to their tenants. Typically this means that tenants must purchase individual VM configurations based on the peak demands of the applications, or be restricted to only scale-out applications that can share a pool of VMs. This diminishes the value proposition of moving to a public cloud as compared to server consolidation in a private virtualized datacenter, where one gets the benefits of statistical multiplexing between VMs belonging to the same or different applications. Ideally one would like to enable a cloud tenant to buy capacity in bulk and benefit from statistical multiplexing among its workloads. This requires the purchased capacity to be dynamically and transparently allocated among the tenant's VMs that may be running on different servers, even across datacenters. In this paper, we propose two novel algorithms called BPX and DBS that are able to provide the cloud customer with the abstraction of buying bulk capacity. These algorithms dynamically allocate the bulk capacity purchased by a customer between its VMs based on their individual demands and user-set importance. Our algorithms are highly scalable and are designed to work in a large-scale distributed environment. We implemented a prototype of BPX as part of VMware's management software and showed that BPX is able to closely mimic the behavior of a centralized allocator in a distributed manner.

Supporting QoS control mechanisms in shared storage arrays is constrained by the well-justified fear of impacting the system efficiency. This motivates our study of the trade off between fairness and efficiency in shared storage systems. We propose two adaptations that can be applied to existing IO scheduling mechanisms: the concurrency bound and the batch size. Although these knobs are well known, their impact on system performance and automatic adaptation based on current workload characteristics have not been studied before. Using synthetic benchmarks and trace workloads, we show that the adaptive proportional share algorithm achieves over 90% IO efficiency while maintaining the specified QoS requirements.

With the growing adoption of virtualized datacenters and cloud hosting services, the allocation and sizing of resources such as CPU, memory, and I/O bandwidth for virtual machines (VMs) is becoming increasingly important. Accurate performance modeling of an application would help users in better VM sizing, thus reducing costs. It can also benefit cloud service providers who can offer a new charging model based on the VMs' performance instead of their configured sizes. In this paper, we present techniques to model the performance of a VM-hosted application as a function of the resources allocated to the VM and the resource contention it experiences. To address this multi-dimensional modeling problem, we propose and refine the use of two machine learning techniques: artificial neural network (ANN) and support vector machine (SVM). We evaluate these modeling techniques using five virtualized applications from the RUBiS and Filebench suite of benchmarks and demonstrate that their median and 90th percentile prediction errors are within 4.36% and 29.17% respectively. These results are substantially better than regression based approaches as well as direct applications of machine learning techniques without our refinements. We also present a simple and effective approach to VM sizing and empirically demonstrate that it can deliver optimal results for 65% of the sizing problems that we studied and produces close-to-optimal sizes for the remaining 35%.

Virtualized datacenters strive to reduce costs through workload consolidation. Workloads exhibit a diverse set of IO behaviors and varying IO load that makes it difficult to estimate the IO performance on shared storage. As a result, system administrators often resort to gross overprovisioning or static partitioning of storage to meet application demands. In this paper, we introduce Pesto, a unified storage performance management system for heterogeneous virtualized datacenters. Pesto is the first system that completely automates storage performance management for virtualized datacenters, providing IO load balancing with cost-benefit analysis, per-device congestion management, and initial placement of new workloads.

Storage management in virtualized environments is considered as one of the biggest cost factors. According to some estimates, majority of the cost and performance problems are related to storage devices. In this tutorial, we will discuss some of the key storage technologies deployed in virtual datacenters. We will discuss a set of unique challenges faced by administrators and users due to increasing number of layers of abstraction. Next we will discuss some tools and techniques to do workload characterization and monitor devices in order to understand and trouble-shoot IO problems. We will also present some of the recent solutions proposed by industry and academia to handle these problems followed by upcoming technological trends and directions for future research.

Increasing power costs has renewed interest in software based power optimizations. One of the key question that any software based mechanism needs to decide is: How long should I wait before changing the power state of a device? In this paper, we study the solution space of this problem based on the workload history and device characteristics. We model the workload as a distribution of idle times (collected online) and the device characteristics using break-even time. Our analysis shows that one should not do any power savings if the mean of the distribution is lower than break even time. We also show that being greedy, (i.e. wait time = 0) gives the highest power savings for most of the well known distributions. Our model is also able to capture special cases such as bimodal distributions.

We examine the problem of providing fair bandwidth allocation in distributed storage systems similar to the research prototypes HP FAB and IBM Intelligent Bricks. The problem poses significant new challenges beyond those encountered in network and storage QoS scheduling. Specifically, resources are heterogeneous i.e. an IO request can only be serviced by a particular server, and service is both requested and provided in a distributed manner with no centralized controller. We provide a distributed algorithm, d-Clock, that runs locally on each of the servers and provides global fairness guarantees without causing resource-specific starvation, with minimal synchronization overhead.

Storage consolidation is becoming an attractive paradigm for data organization because of the economies of sharing and the ease of centralized management. However, sharing of resources is viable only if applications can be isolated from each other. This work targets the problem of providing performance guarantees to an application irrespective of the behavior of other workloads. Application requirements are represented in terms of the average throughput, latency and maximum burst size. Most earlier schemes only do weighted bandwidth allocation; schemes that provide control of latency either cannot handle bursts or penalize applications for their own prior behavior, such as using spare capacity.

High-end shared storage systems serving multiple independent workloads must assure that concurrently executing clients will receive a fair or agreed-upon share of system I/O resources. In a parallel I/O system an application makes requests for specific disks at different steps of its computation depending on the data layout and its computational state. Different applications contend for disk access making the problem of maintaining fair allocation challenging.We propose a model for differentiated disk bandwidth allocation based on lexicographic minimization, and provide new efficient scheduling algorithms to allocate the I/O bandwidth fairly among contending applications. A major contribution of our model is its ability to handle multiple parallel disks and contention for disks among the concurrent applications. Analysis and simulation-based evaluation shows that our algorithms provide performance isolation, weighted allocation of resources, and are work conserving. The solutions are also applicable to other shared resource environments dealing with non-uniform heterogeneous servers.