Abstract — Today’s shared hosting platforms often employ virtualization to allow multiple enterprise applications with time-varying resource demands to share a common infrastructure in order to improve resource utilization. Meeting application-level quality of service (QoS) goals becomes a challenge in such an environment as enterprise applications often have a multi-tier architecture and complex interactions and dependencies among individual tiers. In addition, when the shared infrastructure becomes overloaded, appropriate resource control needs to be performed at these individual tiers in a coordinated fashion in order to provide differentiated services to co-hosted applications. In this paper, we present an adaptive multivariate controller that dynamically adjusts the resource shares to individual tiers of multiple applications in order to meet a specified level of service differentiation. The controller parameters are automatically tuned at runtime based on a quadratic cost function and a system model that is learned online using a recursive least-squares (RLS) method. To evaluate our controller design, we built a testbed hosting two instances of the RUBiS application, a multi-tier online auction web site, using Xen virtual machines. Our results indicate that our controller is able to meet given QoS differentiation targets between co-hosted applications while the total demand from these applications exceeds the capacities of the shared systems. I.

Elasticity—where systems acquire and release resources in response to dynamic workloads, while paying only for what they need—is a driving property of cloud computing. At the core of any elastic system is an automated controller. This paper addresses elastic control for multi-tier application services that allocate and release resources in discrete units, such as virtual server instances of predetermined sizes. It focuses on elastic control of the storage tier, in which adding or removing a storage node or “brick ” requires rebalancing stored data across the nodes. The storage tier presents new challenges for elastic control: actuator delays (lag) due to rebalancing, interference with applications and sensor measurements, and the need to synchronize the multiple control elements, including rebalancing. We have designed and implemented a new controller for elastic storage systems to address these challenges. Using a popular distributed storage system—the Hadoop Distributed File System (HDFS)—under dynamic Web 2.0 workloads, we show how the controller adapts to workload changes to maintain performance objectives efficiently in a pay-as-you-go cloud computing environment.

This paper introduces Automat, a testbed architecture and prototype for research in autonomic services and hosting centers. Automat is an interactive web-based laboratory in which users allocate resources from an ondemand server cluster to experiment with controller policies for sense-and-respond monitoring and adaptation by hosted services and, crucially, by the hosting center itself. Users may explore the interactions of components selected from a menu of applications, workloads, faultloads, and controllers, or install their own components. Components may include view extensions as plugins to a Web portal interface, enabling users to monitor and interact with controllers during an experiment. 1

... (MMOGs) can include millions of concurrent players spread across the world. To keep these highly-interactive virtual environments online, a MMOG operator may need to provision tens of thousands of computing resources from various data centers. Faced with large resource demand variability, and with misfit resource renting policies, the current industry practice is to maintain for each game tens of self-owned data centers. In this work we investigate the dynamic resource provisioning from external data centers for MMOG operation. We introduce a novel MMOG workload model that represents the dynamics of both the player population and the player interactions. We evaluate several algorithms, including a novel neural network predictor, for predicting the resource demand. Using trace-based simulation, we evaluate the impact of the data center policies on the resource provisioning efficiency; we show that dynamic provisioning can be much more efficient than its static alternative.

We introduce a novel multi-resource allocator to dynamically allocate resources for database servers running on virtual storage. Multi-resource allocation involves proportioning the database and storage server caches, and the storage bandwidth between applications according to overall performance goals. The problem is challenging due to the interplay between different resources, e.g., changing any cache quota affects the access pattern at the cache/disk levels below it in the storage hierarchy. We use a combination of on-line modeling and sampling to arrive at near-optimal configurations within minutes. The key idea is to incorporate access tracking and known resource dependencies e.g., due to cache replacement policies, into our performance model. In our experimental evaluation, we use both microbenchmarks and the industry standard benchmarks TPC-W and TPC-C. We show that our multi-resource allocation approach improves application performance by up to factors of 2.9 and 2.4 compared to state-of-the-art singleresource controllers, and their ad-hoc combination, respectively. 1

Creating good adaptation policies is critical to building complex autonomic systems since it is such policies that define the system configuration used in any given situation. While online approaches based on control theory and rulebased expert systems are possible solutions, each has its disadvantages. Here, a hybrid approach is described that uses modeling and optimization offline to generate suitable configurations, which are then encoded as policies that are used at runtime. The approach is demonstrated on the problem of providing dynamic management in virtualized consolidated server environments that host multiple multi-tier applications. Contributions include layered queuing models for Xen-based virtual machine environments, a novel optimization technique that uses a combination of bin packing and gradient search, and experimental results that show that automatic offline policy generation is viable and can be accurate even with modest computational effort. 1

Virtual machines have emerged as an attractive approach for utility computing platforms because applications running on VMs are fault- and security- isolated from each other, yet can share physical machines. An important property of a virtualized utility computing platform is how quickly it can react to changing demand. We refer to the capability of a utility computing platform to quickly reassign resources as the agility of the platform. We are targeting hosting utility provider environments where the entire platform is under the control of a single administrative domain and application instances often form application-level clusters. In this work, we examine resource reassignment mechanisms in these environments from the agility perspective and outline a new mechanism that exploits properties of a virtualized utility computing platform. This new mechanism employs ghost virtual machines (VMs), which participate in application clusters, but do not handle client requests until activated by the resource management system. We evaluate this, as well as other, mechanisms on a utility computing testbed. The results show that this ghost VM approach is superior to other approaches in its agility, and allows a new VM to be added to an existing application cluster in a few seconds with negligible overhead. This is a promising result as we develop resource management algorithms for a globally distributed utility computing platform.

Abstract. Emerging deadline-driven Grid applications require a number of computing resources to be available over a time frame, starting at a specific time in the future. To enable these applications, it is important to predict the resource availability and utilise this information during provisioning because it affects their performance. It is impractical to request the availability information upon the scheduling of every job due to communication overhead. However, existing work has not considered how the precision of availability information influences the provisioning. As a result, limitations exist in developing advanced resource provisioning and scheduling mechanisms. This work investigates how the precision of availability information affects resource provisioning in multiple site environments. Performance evaluation is conducted considering both multiple scheduling policies in resource providers and multiple provisioning policies in brokers, while varying the precision of availability information. Experimental results show that it is possible to avoid requesting availability information for every Grid job scheduled thus reducing the communication overhead. They also demonstrate that multiple resource partition policies improve the slowdown of Grid jobs. 1

Today’s data centers face two critical challenges. First, various customers need to be assured by meeting their required service-level agreements such as response time and throughput. Second, server power consumption must be controlled in order to avoid failures caused by power capacity overload or system overheating due to increasing high server density. However, existing work controls power and application-level performance separately, and thus, cannot simultaneously provide explicit guarantees on both. In addition, as power and performance control strategies may come from different hardware/software vendors and coexist at different layers, it is more feasible to coordinate various strategies to achieve the desired control objectives than relying on a single centralized control strategy. This paper proposes Co-Con, a novel cluster-level control architecture that coordinates individual power and performance control loops for virtualized server clusters. To emulate the current practice in data centers, the power control loop changes hardware power states with no regard to the application-level performance. The performance control loop is then designed for each virtual machine to achieve the desired performance even when the system model varies significantly due to the impact of power control. Co-Con configures the two control loops rigorously, based on feedback control theory, for theoretically guaranteed control accuracy and system stability. Empirical results on a physical testbed demonstrate that Co-Con can simultaneously provide effective control on both application-level performance and underlying power consumption.

Many of today’s applications are delivered as scalable, multi-tier services deployed in large data centers. These services frequently leverage shared, scale-out, key-value storage layers that can deliver low latency under light workloads, but may exhibit significant queuing delay and even dropped requests under high load. Stout is a system that helps these applications adapt to variation in storage-layer performance by treating scalable key-value storage as a shared resource requiring congestion control. Under light workloads, applications using Stout send requests to the store immediately, minimizing delay. Under heavy workloads, Stout automatically batches the application’s requests together before sending them to the store, resulting in higher throughput and preventing queuing delay. We show experimentally that Stout’s adaptation algorithm converges to an appropriate batch size for workloads that require the batch size to vary by over two orders of magnitude. Compared to a non-adaptive strategy optimized for throughput, Stout delivers over 34 × lower latency under light workloads; compared to a non-adaptive strategy optimized for latency, Stout can scale to over 3 × as many requests. 1.