Applications with a dynamic workload demand need access to a flexible infrastructure to meet performance guarantees and minimize resource costs. While cloud computing provides the elasticity to scale the infrastructure on demand, cloud service providers lack control and visibility of user space applications, making it difficult to accurately scale the infrastructure. Thus, the burden of scaling falls on the user. That is, the user must determine when to trigger scaling and how much to scale. Scaling becomes even more challenging when applications exhibit dynamic changes in their behavior. In this paper, we propose a new cloud service, Dependable Compute Cloud (DC2), that automatically scales the infrastructure to meet the user-specified performance requirements, even when multiple user requests execute concurrently. DC2 employs Kalman filtering to automatically learn the (possibly changing) system parameters for each application, allowing it to proactively scale the infrastructure to meet performance guarantees. Importantly, DC2 is designed for the cloud - it is application-agnostic and does not require any offline application profiling or benchmarking, training data, or expert knowledge about the application. We evaluate DC2 via implementation on OpenStack using a multi-tier application under a range of workload mixes and arrival traces. Our experimental results demonstrate the robustness and superiority of DC2 over existing rule-based approaches with respect to avoiding SLA violations and minimizing resource consumption.

Unforeseen events such as node failures and resource contention can have a severe impact on the performance of data processing frameworks, such as Hadoop, especially in cloud environments where such incidents are common. SLA compliance in the presence of such events requires the ability to quickly and dynamically resize infrastructure resources. Unfortunately, the distributed and stateful nature of data processing frameworks makes it challenging to accurately scale the system at run-time. In this paper, we present the design and implementation of a model-driven autoscaling solution for Hadoop clusters. We first develop novel gray-box performance models for Hadoop workloads that specifically relate job execution times to resource allocation and workload parameters. We then employ these models to dynamically determine the resources required to successfully complete the Hadoop jobs as per the user-specified SLA under various scenarios including node failures and multi-job executions. Our experimental results on three different Hadoop cloud clusters and across different workloads demonstrate the efficacy of our models and highlight their autoscaling capabilities.

Infrastructure-as-a-service clouds are becoming widely adopted. However, resource sharing and multi-tenancy have made performance anomalies a top concern for users. Timely debugging those anomalies is paramount for minimizing the performance penalty for users. Unfortunately, this debugging often takes a long time due to the inherent complexity and sharing nature of cloud infrastructures. When an application experiences a performance anomaly, it is important to distinguish between faults with a global impact and faults with a local impact as the diagnosis and recovery steps forfaults with a global impact or local impact are quite different. In this paper, we present PerfCompass, an online performance anomaly fault debugging tool that can quantify whether a production-run performance anomaly has a global impact or local impact. PerfCompass can use this information to suggest the root cause as either an external fault (e.g., environment-based) or an internal fault (e.g., software bugs). Furthermore, PerfCompass can identify top affected system calls to provide useful diagnostic hints for detailed performance debugging. PerfCompass does not require source code or runtime application instrumentation, which makes it practical for production systems. We have tested PerfCompass by running five common open source systems (e.g., Apache, MySQL, Tomcat, Hadoop, Cassandra) inside a virtualized cloud testbed. Our experiments use a range of common infrastructure sharing issues and real software bugs. The results show that PerfCompass accurately classifies 23 out of the 24 tested cases without calibration and achieves 100 percent accuracy with calibration. PerfCompass provides useful diagnosis hints within several minutes and imposes negligible runtime overhead to the production system during normal execution time.

As the cloud services journey through their lifecycle towards becoming commodities, the demand is increasing for "pay-per-use" pricing model. In this model, users are charged for the amount of resources, e.g., Volume of transactions, CPU usage, etc., being consumed during a given time period. Software as a Service (SaaS) providers charging their customers via pay-per-use (e.g., Microsoft Azure Web Services) and facing Infrastructure as a Service (IaaS) costs per VM per month (e.g., Soft Layer) have to carefully choose and scale their non-revenue generating service management infrastructure to penetrate and stay in the market. In this paper, we focus on the metering and rating aspects of cloud service management, and their scalability with the SaaS business and operational changes. We design a framework for cloud service providers to scale their revenue management systems in a cost-aware manner, where the deployment of these revenue systems dynamically uses existing or newly provisioned SaaS VMs, instead of the extant approach of using dedicated setups. Our experimental analysis shows that service management related tasks can be offloaded to the existing VMs with at most 15% overhead in CPU utilization, 10% overhead for memory usage, and negligible overhead for I/O and network usage. We used traces from IBM production servers to mimic the load on VMs. By dynamically scaling the service management setup, we were able to adapt to increasing metering data processing requirements without incurring additional cost, while preserving the infrastructure footprint.

A critical feature of IaaS cloud computing is the ability to quickly disseminate the content of a shared dataset at large scale. In this context, a common pattern is collective on-demand read, i.e., accessing the same VM image or dataset from a large number of V Minstances concurrently. There are various techniques that avoid I/Ocontention to the storage service where the dataset is located without relying on pre-broadcast. Most such techniques employ peer-to-peer collaborative behavior where the VM instances exchange information about the content that was accessed during runtime, such that it impossible to fetch the missing data pieces directly from each other rather than the storage system. However, such techniques are often limited within a group that performs a collective read. In light of high data redundancy on large IaaS data centers and multiple users that simultaneously run VM instance groups that perform collective reads, an important opportunity arises: enabling unrelated VMinstances belonging to different groups to collaborate and exchange common data in order to further reduce the I/O pressure on the storage system. This paper deals with the challenges posed by such absolution, which prompt the need for novel techniques to efficiently detect and leverage common data pieces across groups. To this end, we introduce a low-overhead fingerprint based approach that we evaluate and demonstrate to be efficient in practice for a representative scenario on dozens of nodes and a variety of group configurations.

As the cloud services journey through their life-cycle towards commodities, cloud service providers have to carefully choose the metering and rating tools and scale their infrastructure to effectively process the collected metering data. In this paper, we focus on the metering and rating aspects of the revenue management and their adaptability to business and operational changes. We design a framework for IT cloud service providers to scale their revenue systems in a cost-aware manner. The main idea is to dynamically use existing or newly provisioned SaaS VMs, instead of dedicated setups, for deploying the revenue management systems. At on-boarding of new customers, our framework performs off-line analysis to recommend appropriate revenue tools and their scalable distribution by predicting the need for resources based on historical usage. This allows the revenue management to adapt to the ever evolving business context. We evaluated our framework on a test bed of 20 physical machines that were used to deploy 12 VMs within Open Stack environment. Our analysis shows that service management related tasks can be offloaded to the existing VMs with at most 15% overhead in CPU utilization, 10% overhead for memory usage, and negligible overhead for I/O and network usage. By dynamically scaling the setup, we were able to reduce the metering data processing time by many folds without incurring any additional cost.

IaaS clouds are becoming a standard way of providing elastic compute capacity at an affordable cost. To achieve that, VM provisioning system has to optimally allocate I/O and compute resources. One of the significant optimization opportunities is to leverage content similarity across VM images. While many studies have been devoted to de-duplication of VM images, this paper is to the best of our knowledge, the first one to comprehensively study the relationship between the VM image similarity structure and the runtime I/O access patterns. Our study focuses on block-level similarity and I/O access patterns, revealing correlations between common content of different images and application-level access semantics. Furthermore, it also zooms on several runtime I/O access pattern aspects, such as the similarity of the sequences in which common content is accessed. The results show a strong tendency for access pattern locality within common content clusters across VM images, regardless of the rest of the composition. Furthermore, it reveals a strong tendency for read accesses to refer to the same subsets of blocks within the common content clusters, while preserving the same ordering. These results provide important insights that can be used to optimize on-demand VM image content delivery under concurrency.

Cloud computing offers the flexibility to dynamically size the infrastructure in response to changes in workload demand. While both horizontal and vertical scaling of infrastructure is supported by major cloud providers, these scaling options differ significantly in terms of their cost, provisioning time, and their impact on workload performance. Importantly, the efficacy of horizontal and vertical scaling critically depends on the workload characteristics, such as the workload's parallelizability and its core scalability. In today's cloud systems, the scaling decision is left to the users, requiring them to fully understand the tradeoffs associated with the different scaling options. In this paper, we present our solution for optimizing the resource scaling of cloud deployments via implementation in OpenStack. The key component of our solution is the modelling engine that characterizes the workload and then quantitatively evaluates different scaling options for that workload. Our modelling engine leverages Amdahl's Law to model service time scaling in scaleup environments and queueing-theoretic concepts to model performance scaling in scale-out environments. We further employ Kalman filtering to account for inaccuracies in the model-based methodology, and to dynamically track changes in the workload and cloud environment.

With multiplicity of Cloud service providers offering geographically distributed Compute Clouds, both providers and customers find it necessary to quickly move virtual machine images between data centers. This is usually accomplished using standard rsync-based transfer which is slow and bandwidth intensive given large size of virtual machine images. This article proposes a mechanism to reconstitute an image on target data center using information about overlap among images and content already available in the target data center. No changes to file system are needed and the approach can immediately be used with traditional libraries storing images as regular files. Moreover, we use a peer-to-peer approach that allows simultaneous retrieval of fragments from multiple data centers. The system and algorithms have been implemented and evaluated on two Compute Cloud environments using three image libraries representative for a typical service provider. The evaluation shows an average 6 times reduction in terms of network transfer volume and time and can result in even larger reductions in case of images with small configuration changes.

Virtualized data centers became ubiquitous and led to invention of new delivery model called Cloud computing. Cloud service providers pursue high degree of automation for the management processes in order to make their services responsive and inexpensive. One of such critical processes is virtual machine provisioning. This article proposes and evaluates a provisioning model that leverages virtual machine image similarity to reduce the data volume transferred from the storage server to the hypervisor on which the virtual machine is being instantiated. We also present an analytical model of such a provisioning process for a single hypervisor shedding light on the impact of degree of image similarity, system utilization, and hypervisor capacity on the performance of the system. The model is validated using discrete event simulator. The algorithm has been implemented on a testbed system and also extensive simulations of virtualized server clusters were conducted. The proposed provisioning scheme can offer significant (up to 80%) reduction in data transfer between storage server and hypervisors thus significantly reducing provisioning time while also decreasing cost.

Virtualized data centers became ubiquitous and led to invention of new delivery model called cloud computing. Cloud service providers pursue high degree of automation for the management processes in order to make their services responsive and inexpensive. One of such critical processes is virtual machine provisioning. This article reports on our on-going research to evaluate the efficiency of the provisioning process that leverages virtual machine similarity to reduce the amount of data that has to be transferred from the storage server to the hyper visor on which the virtual machine is being instantiated. The simulation experiments point to conclusion that the redundancy based provisioning can provide very substantial reduction in terms of amount of data that needs to be copied from the storage server.

Virtualization has rapidly gained popularity affecting multiple levels of computing stack. Since it decouples resources from their users it provides greater flexibility in terms of resource allocation but also brings new challenges for optimal design, provisioning and runtime management of systems. Cloud computing is a paradigm of computing that offers virtualized resources “as a service.” Cloud Managers are responsible for lifecycle management of virtual resources, efficient utilization of physical resources and exposing basic operational APIs to users. Software solutions can then be deployed on these virtual resources. In this paper we propose a Virtual Hypervisor abstraction allowing solution managers to have an improved control over the resource allocation decisions regarding their virtual machines while maintaining cloud manager's role as the ultimate physical resource manager. We also introduce a novel resource allocation algorithm illustrating how the Virtual Hypervisor abstraction can be efficiently realized by the global cloud manager. We also use simulations to illustrate that our algorithm can be used to achieve fairer resource sharing and isolation across different Virtual Hypervisors.

Desktop virtualization is a new computing approach to application delivery and management. It leverages OS virtualization and remoting protocols to provide users with remote access to virtual machines running in a centralized data center. It promises significant benefits in terms of improved data security, reduced management complexity, and more efficient and flexible resource usage. However, it brings a lot of management challenges typical for centralized systems, with performance and quality of service management being one of the most important. This paper proposes a management algorithm suitable for efficient resource allocation in virtualized desktop environments and takes application performance QoS features into consideration. It proposes a novel, non-intrusive method for application and remoting protocol agnostic desktop responsiveness monitoring. Moreover, it is based on studies of desktop workload usage which enabled us to discover and leverage workload patterns that can lead to increased efficiency both in terms of desktop responsiveness and resource usage. We have prototyped the system and discuss several case studies validating the approach and illustrating the most important features of the solution.

The thin-client computing model has been recently regaining popularity in a new form known as the virtual desktop. That is where the desktop is hosted on a virtualized platform. Even though the interest in this computing paradigm is broad there are relatively few tools and methods for benchmarking virtual client infrastructures. We believe that developing such tools and approaches is crucial for the future success of virtual client deployments and also for objective evaluation of existing and new algorithms, communication protocols, and technologies. We present DeskBench, a virtual desktop benchmarking tool, that allows for fast and easy creation of benchmarks by simple recording of the user's activity. It also allows for replaying the recorded actions in a synchronized manner at maximum possible speeds without compromising the correctness of the replay. The proposed approach relies only on the basic primitives of mouse and keyboard events as well as screen region updates which are common in window manager systems. We have implemented a prototype of the system and also conducted a series of experiments measuring responsiveness of virtual machine based desktops under various load conditions and network latencies. The experiments illustrate the flexibility and accuracy of the proposed method and also give some interesting insights into the scalability of virtual machine based desktops.

Traditional desktop delivery model is based on a large number of distributed PCs executing operating system and desktop applications. Managing traditional desktop environments is incredibly challenging and costly. Tasks like installations, configuration changes, security measures require time-consuming procedures and dedicated deskside support. Also these distributed desktops are typically underutilized, resulting in low ROI for these assets. Further, this distributed computing model for desktops also creates a security concern as sensitive information could be compromised with stolen laptops or PCs. Desktop virtualization, which moves computation to the data center, allows users to access their applications and data using stateless ldquothin-clientldquo devices and therefore alleviates some of the problems of traditional desktop computing. Enterprises can now leverage the flexibility and cost-benefits of running users' desktops on virtual machines hosted at the data center to enhance business agility and reduce business risks, while lowering TCO. Recent research and development of cloud computing paradigm opens new possibilities of mass hosting of desktops and providing them as a service. However, transformation of legacy systems to desktop clouds as well as proper capacity provisioning is a challenging problem. Desktop cloud needs to be appropriately designed and provisioned to offer low response time and good working experience to desktop users while optimizing back-end resource usage and therefore minimizing provider's costs. This paper presents tools and approaches we have developed to facilitate fast and accurate planning for desktop clouds. We present desktop workload profiling and benchmarking tools as well as desktop to cloud transformation process enabling fast and accurate transition of legacy systems to new cloud-based model.

Desktop virtualization is a new delivery method in which desktop operating systems execute in a data center and users access their applications using stateless ¿thin-client¿ devices. This paradigm promises significant benefits in terms of data security, flexibility, and reduction of the total cost of ownership. However, in order to further optimize this approach while maintaining good user experience, efficient resource management algorithms are required. This paper formulates an analytical model allowing for detailed investigation of how power consumption of virtualized server farm depends on properties of workload, adaptiveness of virtualization infrastructure, and average density of virtual machines per physical server. Assumptions needed to develop the model are confirmed using statistical analysis of desktop workload traces and the model itself is also validated using simulations. We apply the model to compare power consumption of static and dynamic virtual machine allocation strategies. The results of the study can be used to develop online virtual machine migration algorithms. Even though this paper focuses on virtualized systems running desktop workloads, the modeling approach is general and can be applied in other contexts.

A dynamic server migration and consolidation algorithm is introduced. The algorithm is shown to provide substantial improvement over static server consolidation in reducing the amount of required capacity and the rate of service level agreement violations. Benefits accrue for workloads that are variable and can be forecast over intervals shorter than the time scale of demand variability. The management algorithm reduces the amount of physical capacity required to support a specified rate of SLA violations for a given workload by as much as 50% as compared to static consolidation approach. Another result is that the rate of SLA violations at fixed capacity may be reduced by up to 20%. The results are based on hundreds of production workload traces across a variety of operating systems, applications, and industries.

Virtualization rapidly gains popularity affecting multiple levels of the computing stack. Since it decouples resources from their users it provides greater flexibility in terms of resource allocation but also brings new challenges for optimal design, provisioning, and runtime management of systems. This paper presents an analytical model of virtual machine migration that provides an estimate of the expected improvement in response time due to a migration decision. The model explicitly accounts for predictability of resource demands and characteristics of the virtualization infrastructure, such as migration time and overhead. It brings new insights into the decision strategies that should be used to improve the overall responsiveness of the system when dynamically managing virtual machines. The model is validated using simulations driven by data center resource utilization traces.

Timestep stochastic simulation (TSS) is a novel method for generating sample paths of computer networks, with low computation cost independent of packet rates. It has accuracy adequate to evaluate general network and flow configurations, including arbitrary flow start times and durations, drop-tail queuing (i.e., does not require RED), and arbitrary state-dependent control mechanisms for congestion control and routing. TSS generates the evolution of the system state S(t) on a sample path in time steps of size \delta. At each step, S(t+\delta) is randomly chosen according to S(t) and the probability distribution Pr[S(t+\delta|S(t)] obtained using the diffusion approximation. Because packet transmission and reception events are replaced by time steps, TSS generates sample paths at a fraction of the cost of packet-level simulation. Because TSS generates sample paths, control feedback can be based on sample path metrics, rather than ensemble metrics, thereby accurately capturing the effects of state-dependent control mechanisms.

As businesses have grown, so has the need to deploy I/T applications rapidly to support the expanding business processes. Often, this growth was achieved in an unplanned way: each time a new application was needed a new server along with the application software was deployed and new storage elements were purchased. In many cases this has led to what is often referred to as "server sprawl", resulting in low server utilization and high system management costs. An architectural approach that is becoming increasingly popular to address this problem is known as server virtualization. In this paper we introduce the concept of server consolidation using virtualization and point out associated issues that arise in the area of application performance. We show how some of these problems can be solved by monitoring key performance metrics and using the data to trigger migration of virtual machines within physical servers. The algorithms we present attempt to minimize the cost of migration and maintain acceptable application performance levels

Physical layer capture (PLC) in 802.11b refers to the successful reception of the stronger (higher signal strength at receiver) frame in a collision. PLC causes significant imbalance in the throughputs of sources. Existing 802.11b simulators, including ns2 and Qualnet, assume that PLC occurs only if the stronger frame arrives first at the receiver. We show empirically that in reality PLC occurs even if the stronger frame arrives later (but within the physical layer preamble of the first frame). Consequently, throughput unfairness in reality can be significantly (up to 15%) higher than with the former PLC model. We have modified the ns2 simulator to account for this and Qualnet is incorporating a fix in their next release. To identify which frames were involved in collisions, when their transmissions started, and which of them were retrieved, we have devised a novel technique using multiple sniffers and instrumented device drivers to reconstruct from the air interface all tx/rx events in a WLAN to within 4 /spl mu/s accuracy. This allows us to quantify the causal links from the PHY layer through the MAC layer to the observed application layer imbalance. It also shows that the arrival times of colliding frames routinely differ by as much as 20 /spl mu/s due to inherent uncertainties of 802.11b firmware clock synchronization and rx/tx turnaround delays, and that the frame to arrive first can be either the stronger or the weaker with equal likelihood.

This paper explores merging SAN performance management tools into the host system thus dissolving the traditional boundary between hosts and storage. The host service time distribution is exploited to locate caches in the SAN, detect and isolate contention in the communication channels and backend disk drives. We define essential metrics and measure I/O workloads over extended periods to define baseline behavior. We have implemented architecture to acquire configuration and performance data from agents on the host and SAN elements. In addition to performance management this data is the basis of management functions such as capacity planning, and sizing host and SAN resources required for application rollout and expansion. The static dependency map (SAN topology or 'wiring' diagram), provided by current SAN management tools, is extended into the host server. We then attempt to develop a corresponding dynamic dependency view showing how performance metrics at the application and host level, such as I/O queuing levels and service times, relate to response times in the SAN. The dynamic view is also an important complement to the static dependency graph because the actual traffic flow may be difficult to deduce from the topology.