Edge computing is a promising paradigm for deploying latency-sensitive applications (Apps) as it brings resources closer to end users. Edge Apps often adopt a microservice (MS) architecture, breaking monolithic Apps into lightweight, containerized MSs that can be dynamically and independently deployed. However, managing such Apps involves three key challenges: (i) optimizing the placement of MSs to reduce both response time and resource overhead, (ii) handling MS migration or relocation as users move while minimizing App service disruption (App downtime), and (iii) enabling MS sharing across Apps while ensuring performance guarantees. We formulate this as an optimization problem, named Multi-microservice Application Placement (MAP), prove its NP-hardness, and introduce STEP (State and Topology-aware Edge-MS Placement), a polynomial-time heuristic. STEP distinguishes itself from prior work by: (i) jointly considering stateful and stateless MS characteristics in deployment decisions, (ii) exploiting MS shareability to reduce resource usage, (iii) balancing response latency, App downtime, and resource utilization, and (iv) leveraging multiple versions of the same MS to adapt quality of service to available edge resources. Our results in a small-scale scenario show that STEP achieves near-optimal performance with only 7% higher CPU cost than the optimal solution. Large-scale real-time experiments on a Kubernetes cluster demonstrate that STEP consistently outperforms competing methods, achieving up to 50% lower deployment costs while delivering 50% gain in app quality and saving 15% in radio resources with over 90% request success rates.

Containers have become the standard for deploying applications in many cloud systems due to its convenience. However, this convenience leads to significant container bloat, i.e., unused files that inflate container image sizes, increase provisioning times, waste resources and introduce security vulnerabilities. Bloat is particularly problematic in serverless and edge computing scenarios, where resources are constrained, and performance is critical, and for microservice applications where rapid scaling is key to meet performance targets. However, existing container debloating tools are often limited in both effectiveness and robustness. In this paper, we propose BLAFS, a bloat-aware container filesystem that removes bloat while guaranteeing the correct operation of the debloated containers. BLAFS addresses bloat at the filesystem level by introducing new layers in the filesystem to enable debloating. During runtime, accessed files are moved to the debloating layers, and then similar to garbage collection mechanisms, BLAFS removes files that are not accessed during runtime. An optional reloading layer fetches files from a remote cloud cache on-demand if the files are mistakenly removed. We discuss how BLAFS can be used in different deployment scenarios and for different use-cases including container security-hardened and a dynamic deployment mode where the target is improved provisioning performance. We evaluate BLAFS performance using the top 20 downloaded containers from DockerHub, four ML containers, and SEBS, a Serverless Benchmark containing 10 serverless functions and compare its performance against two state-of-the-art debloating tools. Our evaluation shows that BLAFS reduces container sizes by up to 95% and cold-starts by up to 68%. In the security-hardened mode, BLAFS removes up to 89% of CVEs while the two state-of-the-art debloating tools fail on most of the workloads. We identify their limitations, and show how BLAFS provides a more principled approach to debloating. Additionally, when combined with lazy-loading snapshotters, BLAFS improves provisioning efficiency, reducing conversion times by up to 93% and provisioning times by up to 19%.

Edge computing reduces latency for mobile applications (Apps) by processing data closer to users, while containerized microservices (MSs) enable their modular deployment. Managing such Apps involves three key challenges: (i) strategically placing MSs to minimize response latency and resource consumption, (ii) managing MS migration/relocation during user mobility or traffic load changes while limiting App downtime, and (iii) enabling MS sharing across Apps while ensuring target performance. We formulate this as an optimization problem (proven to be NP-hard) and propose STEP, a polynomial-time heuristic. In contrast to prior art, STEP (i) jointly considers stateful and stateless MSs in its decisions, (ii) exploits MS shareability to reduce resource usage, (iii) balances response latency, App downtime, and resource utilization, and (iv) leverages multiple versions of the same MS to adapt QoS to available edge resources. Results show that STEP achieves near-optimal performance with only 1.6% higher deployment cost while reducing CPU usage by 42% compared to baselines. Also, it enables real-time App deployment in a large-scale scenario on a Kubernetes cluster with sub-second order execution time and reduced deployment cost by 16-17% compared to its benchmarks.

Managing the performance of online web services in cloud data centers while optimizing resource allocation and power consumption is a multifaceted challenge. Often, resource and power management techniques, such as elastic scaling and power capping, are handled independently, leading to conflicts and sub-optimal power-performance trade-offs. To tackle this issue, we introduce SLO-Power, a system that coordinates the resource and power scaling techniques to achieve power savings while adhering to service level objectives (SLOs), such as tail latency constraints. Our approach employs a combination of analytic queuing models and feedback-driven techniques to jointly allocate resources and power to cloud applications in an SLO and power-aware manner. We implement a prototype of our system and evaluate it using realistic workloads to demonstrate its ability to harmonize elastic and power scaling, enabling enhanced resource utilization and reduced power consumption while ensuring the application performance. Our findings indicate that SLO-Power achieves exceptional power and resource efficiency, approaching near-optimal power-efficiency levels at 90%, all while preventing SLO violations. Furthermore, compared to state-of-the-art solutions, SLO-Power demonstrates lower P95 latency, accompanied by a 12% reduction in resource usage.

Reducing tail latency has become a crucial issue for optimizing the performance of online cloud services and distributed applications. In distributed applications, there are many causes of high end-to-end tail latency, including operating system delays, request re-ordering due to fan-out/fanin, and network congestion. Although recent research has focused on reducing tail latency for individual application components, such as by replicating requests and scheduling, in this paper, we argue for a holistic approach for reducing the end-to-end tail latency across application components. We propose TailClipper, a distributed scheduler that tags each arriving request with an arrival timestamp, and propagates it across the microservices' call chain. TailClipper then uses arrival timestamps to implement an oldest request first scheduler that combines global first-come first serve with a limited form of processor sharing to reduce end-to-end tail latency. In doing so, TailClipper can counter the performance degradation caused by request reordering in multi-tiered and microservices-based applications. We implement TailClipper as a userspace Linux scheduler and evaluate it using cloud workload traces and a real-world microservices application. Compared to state-of-the-art schedulers, our experiments reveal that TailClipper improves the 99th percentile response time by up to 81%, while also improving the mean response time and the system throughput by up to 54% and 29% respectively under high loads.

Many edge applications rely on expensive Deep-Neural-Network (DNN) inference-based video analytics. Typically, a single instance of an inference service analyzes multiple realtime camera streams concurrently. In many cases, only a fraction of these streams contain objects-of-interest at a given time. Hence, it is a waste of computational resources to process all frames from all cameras using the DNNs. On-camera filtration of frames has been suggested as a possible solution to improve the system efficiency and reduce resource wastage. However, many cameras do not have on-camera processing or filtering capabilities. In addition, filtration can be enhanced if frames across the different feeds are selected and prioritized for processing based on the system load and the available resource capacity. This paper introduces CVF, a Cross-video Filtration framework designed around video content and resource constraints. The CVF pipeline leverages compressed-domain data from encoded video formats, lightweight binary classification models, and an efficient prioritization algorithm. This enables the effective filtering of cross-camera frames from multiple sources, processing only a fraction of frames using resource-intensive DNN models. Our experiments show that CVF is capable of reducing the overall response time of video analytics pipelines by up to 50% compared to state-of-the-art solutions while increasing the throughput by up to 120%.

Numerous edge applications that rely on video analytics demand precise, low-latency processing of multiple video streams from cameras. When these cameras are mobile, such as when mounted on a car or a robot, the processing load on the shared edge GPU can vary considerably. Provisioning the edge with GPUs for the worst-case load can be expensive and, for many applications, not feasible.

Traditional power reduction techniques such as DVFS or RAPL are challenging to use with web services because they significantly affect the services’ latency and throughput. Previous work suggested the use of controllers based on control theory or machine learning to reduce performance degradation under constrained power. However, generating these controllers is challenging as every web service applications running in a data center requires a power-performance model and a fine-tuned controller. In this paper, we present DDPC, a system for autonomic data-driven controller generation for power-latency management. DDPC automates the process of designing and deploying controllers for dynamic power allocation to manage the power-performance trade-offs for latency-sensitive web applications such as a social network. For each application, DDPC uses system identification techniques to learn an adaptive power-performance model that captures the application’s power-latency trade-offs which is then used to generate and deploy a Proportional-Integral (PI) power controller with gain-scheduling to dynamically manage the power allocation to the server running application using RAPL. We evaluate DDPC with two realistic latency-sensitive web applications under varying load scenarios. Our results show that DDPC is capable of autonomically generating and deploying controllers within a few minutes reducing the active power allocation of a web-server by more than 50% compared to state-of-the-art techniques while maintaining the latency well below the target of the application.

Since emerging edge applications such as Internet of Things (IoT) analytics and augmented reality have tight latency constraints, hardware AI accelerators have been recently proposed to speed up deep neural network (DNN) inference run by these applications. Resource-constrained edge servers and accelerators tend to be multiplexed across multiple IoT applications, introducing the potential for performance interference between latency-sensitive workloads. In this article, we design analytic models to capture the performance of DNN inference workloads on shared edge accelerators, such as GPU and edgeTPU, under different multiplexing and concurrency behaviors. After validating our models using extensive experiments, we use them to design various cluster resource management algorithms to intelligently manage multiple applications on edge accelerators while respecting their latency constraints. We implement a prototype of our system in Kubernetes and show that our system can host 2.3× more DNN applications in heterogeneous multi-tenant edge clusters with no latency violations when compared to traditional knapsack hosting algorithms.

Edge cloud systems reduce the latency between users and applications by offloading computations to a set of small-scale computing resources deployed at the edge of the network. However, since edge resources are constrained, they can become saturated and bottlenecked due to increased load, resulting in an exponential increase in response times or failures. In this paper, we argue that an application can be split between the edge and the cloud, allowing for better performance compared to full migration to the cloud, releasing precious resources at the edge. We model an application's internal call-Graph as a Directed-Acyclic-Graph. We use this model to develop MicroSplit, a tool for efficient splitting of microservices between constrained edge resources and large-scale distant backend clouds. MicroSplit analyzes the dependencies between the microservices of an application, and using the Louvain method for community detection-a popular algorithm from Network Science-decides how to split the microservices between the constrained edge and distant data centers. We test MicroSplit with four microservice based applications in various realistic cloud-edge settings. Our results show that Microsplit migrates up to 60 % of the microservices of an application with a slight increase in the mean-response time compared to running on the edge, and a latency reduction of up to 800 % compared to migrating the entire application to the cloud. Compared to other methods from the State-of-the-Art, MicroSplit reduces the total number of services on the edge by up to five times, with minimal reduction in response times.

While 360° videos are gaining popularity due to the emergence of VR technologies, storing and streaming such videos can incur up to 20X higher overheads than traditional HD content. Edge caching, which involves caching and serving 360° videos from edge servers, is one possible approach for addressing these overheads. Prior work on 360° video caching has been based on using past history to cache tiles that are likely to be in a viewer's field of view and has not considered methods to intelligently share a limited edge cache across a set of videos that exhibit large variations in their popularity, size, content, and user abandonment patterns. Towards this end, we present CAVE, an adaptive edge caching framework that intelligently optimizes cache allocation across a set of videos taking into account video content, size, and popularity. Our experiments using realistic video workloads shows CAVE improves cache hit-rates, and thus network saving, by up to 50% over state-of-the-art approaches, while also scaling to up to two thousand videos per edge cache. In addition, in terms of scalability, our developed algorithm is embarrassingly parallel, allowing CAVE to scale beyond state-of-the-art solutions that typically do not support parallelization.

Serverless computing has emerged as a new paradigm for running short-lived computations in the cloud. Due to its ability to handle IoT workloads, there has been considerable interest in running serverless functions at the edge. However, the constrained nature of the edge and the latency sensitive nature of workloads result in many challenges for serverless platforms. In this paper, we present LaSS, a platform that uses model-driven approaches for running latency-sensitive serverless computations on edge resources. LaSS uses principled queuing-based methods to determine an appropriate allocation for each hosted function and auto-scales the allocated resources in response to workload dynamics. LaSS uses a fair-share allocation approach to guarantee a minimum of allocated resources to each function in the presence of overload. In addition, it utilizes resource reclamation methods based on container deflation and termination to reassign resources from over-provisioned functions to under-provisioned ones. We implement a prototype of our approach on an OpenWhisk serverless edge cluster and conduct a detailed experimental evaluation. Our results show that LaSS can accurately predict the resources needed for serverless functions in the presence of highly dynamic workloads, and reprovision container capacity within hundreds of milliseconds while maintaining fair share allocation guarantees.

Edge computing has emerged as a popular paradigm for running latency-sensitive applications due to its ability to offer lower network latencies to end-users. In this paper, we argue that despite its lower network latency, the resource-constrained nature of the edge can result in higher end-to-end latency, especially at higher utilizations, when compared to cloud data centers. We study this edge performance inversion problem through an analytic comparison of edge and cloud latencies and analyze conditions under which the edge can yield worse performance than the cloud. To verify our analytic results, we conduct a detailed experimental comparison of the edge and the cloud latencies using a realistic application and real cloud workloads. Both our analytical and experimental results show that even at moderate utilizations, the edge queuing delays can offset the benefits of lower network latencies, and even result in performance inversion where running in the cloud would provide superior latencies. We finally discuss practical implications of our results and provide insights into how application designers and service providers should design edge applications and systems to avoid these pitfalls.

Recent years have witnessed sensors becoming an indispensable part of our life with the camera being one of the most popular and widely deployed sensors. The camera gives rise to numerous vision-based IoT applications that generate high-level understandings of a live video stream by performing analysis on end devices like mobile or embedded devices. Typically, these applications are built with deep learning (DL) models to conduct complex vision tasks, e.g., image classification and object detection. Due to the prohibitive cost of running DL models on end devices close to the camera and with limited computation capabilities, it is widely adopted to offload the computation to a nearby powerful edge server. However, there is a gap between the restricted offloading bandwidth of the end device and the large volume of image data incurred by the live video stream. In this paper, we present Deep Contextualized Compressive Offloading for Images (DCCOI), a lightweight, context-aware, and bandwidth-efficient offloading framework for images. DCCOI consists of the spatial-adaptive encoder, a lightweight neural network, to spatial-adaptively compress the image, and the generative decoder for reconstructing the image from the compressed data. In contrast to existing DL-based encoders, the spatial-adaptive encoder allows an image region to be encoded into different numbers of feature values based on the information in it. This offers a variable-length coding method for image compression, which is a more optimal way for compression than the fix-length coding method took by existing DL-based compression approaches and demonstrates superior accuracy-compression rate trade-offs. We evaluate DCCOI against several baseline compression techniques while serving an object detection-based application. The results show that DCCOI roughly reduces the offloading size of JPEG by a factor of 9 and DeepCOD, the state-of-the-art offloading approach, by 20% with similar accuracy and a compression overhead less than 50ms.

The rapid adoption and the diversification of cloud computing technology exacerbate the importance of a sound experimental methodology for this domain. This work investigates how to measure and report performance in the cloud, and how well the cloud research community is already doing it. We propose a set of eight important methodological principles that combine best-practices from nearby fields with concepts applicable only to clouds, and with new ideas about the time-accuracy trade-off. We show how these principles are applicable using a practical use-case experiment. To this end, we analyze the ability of the newly released SPEC Cloud IaaS benchmark to follow the principles, and showcase real-world experimental studies in common cloud environments that meet the principles. Last, we report on a systematic literature review including top conferences and journals in the field, from 2012 to 2017, analyzing if the practice of reporting cloud performance measurements follows the proposed eight principles. Worryingly, this systematic survey and the subsequent two-round human reviews, reveal that few of the published studies follow the eight experimental principles. We conclude that, although these important principles are simple and basic, the cloud community is yet to adopt them broadly to deliver sound measurement of cloud environments.

Transient computing has become popular in public cloud environments for running delay-insensitive batch and data processing applications at low cost. Since transient cloud servers can be revoked at any time by the cloud provider, they are considered unsuitable for running interactive application such as web services. In this paper, we present VM deflation as an alternative mechanism to server preemption for reclaiming resources from transient cloud servers under resource pressure. Using real traces from top-tier cloud providers, we show the feasibility of using VM deflation as a resource reclamation mechanism for interactive applications in public clouds. We show how current hypervisor mechanisms can be used to implement VM deflation and present cluster deflation policies for resource management of transient and on-demand cloud VMs. Experimental evaluation of our deflation system on a Linux cluster shows that microservice-based applications can be deflated by up to 50% with negligible performance overhead. Our cluster-level deflation policies allow overcommitment levels as high as 50%, with less than a 1% decrease in application throughput, and can enable cloud platforms to increase revenue by 30%.

Spatio-temporal action localization of human actions in a video has been a popular topic over the past few years. It tries to localize the bounding boxes, the time span and the class of one action, which summarizes information in the video and helps humans understand it. Though many approaches have been proposed to solve this problem, these efforts have only focused on perspective videos. Unfortunately, perspective videos only cover a small field-of-view (FOV), which limits the capability of action localization. In this paper, we develop a comprehensive approach to real-time spatio-temporal localization that can be used to detect actions in 360 videos. We create two datasets named UCF-101-24-360 and JHMDB-21-360 for our evaluation. Our experiments show that our method consistently outperforms other competing approaches and achieves a real-time processing speed of 15fps for 360 videos.

Power budgeting is a commonly employed solution to reduce the negative consequences of high power consumption of large scale data centers. While various power budgeting techniques and algorithms have been proposed at different levels of data center infrastructures to optimize the power allocation to servers and hosted applications, testing them has been challenging with no available simulation platform that enables such testing for different scenarios and configurations. To facilitate evaluation and comparison of such techniques and algorithms, we introduce a simulation model for Quality-of-Service aware power budgeting and its implementation in CloudSim. We validate the proposed simulation model against a deployment on a real testbed, showcase simulator capabilities, and evaluate its scalability.

Constantly growing power consumption of data centers is a major concern from environmental and economical reasons. Current approaches to reduce negative consequences of high power consumption focus on limiting the peak power consumption. During high workload periods, power consumption of highly utilized servers is throttled to stay within the power budget. However, the peak power reduction affects performance of hosted applications and thus leads to Quality of Service violations. In this paper, we introduce Power Shepherd, a hierarchical system for application performance aware power shifting. Power Shepherd reduces the data center operational costs by redistributing the available power among applications hosted in the cluster. This is achieved by, assigning server power budgets by the cluster controller, enforcing these power budgets using Running Average Power Limit (RAPL), and prioritizing applications within each server by adjusting the CPU scheduling configuration. We implement a prototype of the proposed solution and evaluate it in a real testbed equipped with power meters and using representative cloud applications. Our experiments show that Power Shepherd has potential to manage a cluster consisting of thousands of servers and limit the increase of operational costs by a significant amount when the cluster power budget is limited and the system is overutilized. Finally, we identify some outstanding challenges regarding model sensitivity and the fact that this approach in its current from is not beneficial to be used in all situations, e.g., when the system is underutilized.

Data centers and clouds are increasingly offering low-cost computational resources in the form of transient virtual machines. Whenever demand for computational resources exceeds their availability, transient resources can reclaimed by preempting the transient VMs. Conventionally, these transient VMs are used by low-priority applications that can tolerate the disruption caused by preemptions.

Many cloud providers offer servers with transient availability at a reduced cost. These servers can be unilaterally revoked by the provider, usually after a warning period to the user. Until recently, it has been thought that these servers are not suitable to run latency-sensitive workloads due to their transient availability. In this paper, we introduce SpotWeb, a framework for running latency-sensitive web workloads on transient computing platforms while maintaining the Quality-of-Service (QoS) of the running applications. SpotWeb is based on three novel concepts; using multi-period optimization---a novel approach developed in finance---for server selection; transiency-aware load-balancing; and using intelligent capacity over-provisioning. We implement SpotWeb and evaluate its performance in both simulations and testbed experiments. Our results show that SpotWeb reduces costs by up to 50% compared to state-of-the-art solutions while being scalable to hundreds of cloud server configurations.

Cloud platforms often execute parallel batch applications, such as distributed machine learning (ML), that include numerous synchronization barriers. These barriers, which prevent any task from advancing beyond a specified point until all tasks have reached that point, significantly degrade application performance by reducing it to that of the slowest "straggler" task. To address the problem, researchers have proposed numerous straggler mitigation techniques, including speculatively re-executing straggler tasks and various relaxations of strict barrier semantics. While these techniques improve parallel application performance, they incur a cost in terms of the resources wasted re-executing tasks or waiting. Importantly, these costs, which are often implicit in prior work that targets dedicated resources, become explicit in the cloud, which charges for resources at fine-grained intervals. In addition, the cost difference between techniques is exacerbated in cloud platforms, since they charge substantially less for transient resources that effectively yield a probabilistic performance across a wide range. While transient resources' low list price is attractive, revocations increase the frequency and severity of stragglers, which decreases parallel job performance and increases overall execution cost. To better understand the cost of synchronization, we develop simple analytical models of different straggler mitigation techniques and compare their cost and performance on on-demand and transient resources. Our analysis shows that i) transient servers offer complex tradeoffs compared to on-demand servers, and can result in higher overall costs despite their highly discounted price due to their probabilistic performance; ii) common approaches to straggler mitigation, which is a well-studied problem, are less effective using transient servers that cause frequent and severe stragglers; and iii) a recent approach to flexible synchronization offers the best cost and performance.

Amazon introduced spot instances in December 2009, enabling "customers to bid on unused Amazon EC2 capacity and run those instances for as long as their bid exceeds the current Spot Price.'' Amazon's real-time computational spot market was novel in multiple respects. For example, it was the first (and to date only) large-scale public implementation of market-based resource allocation based on dynamic pricing after decades of research, and it provided users with useful information, control knobs, and options for optimizing the cost of running cloud applications. Spot instances also introduced the concept of transient cloud servers derived from variable idle capacity that cloud platforms could revoke at any time. Transient servers have since become central to efficient resource management of modern clusters and clouds. As a result, Amazon's spot market was the motivation for substantial research over the past decade. Yet, in November 2017, Amazon effectively ended its realtime spot market by announcing that users no longer needed to place bids and that spot prices will "...adjust more gradually, based on longer-term trends in supply and demand.'' The changes made spot instances more similar to the fixed-price transient servers offered by other cloud platforms. Unfortunately, while these changes made spot instances less complex, they eliminated many benefits to sophisticated users in optimizing their applications. This paper provides a retrospective on Amazon's real-time spot market, including its advantages and disadvantages for allocating transient servers compared to current fixed-price approaches. We also discuss some fundamental problems with Amazon's spot market, which we identified in prior work (from 2016), that predicted its eventual end. We then discuss potential options for allocating transient servers that combine the advantages of Amazon's real-time spot market, while also addressing the problems that likely led to its elimination.

Today's cloud data centers host a wide range of applications including data analytics, batch processing, and interactive processing. These applications require high throughput, low latency, and high reliability from the network. Satisfying these requirements in the face of dynamically varying network conditions remains a challenging problem. Multi-Path TCP (MPTCP) is a recently proposed IETF extension to TCP that divides a conventional TCP flow into multiple subflows so as to utilize multiple paths over the network. Despite the theoretical and practical benefits of MPTCP, its effectiveness for cloud applications and environments remains unclear as there has been little work to quantify the benefits of MPTCP for real cloud applications. We present a broad empirical study of the effectiveness and feasibility of MPTCP for data center and cloud applications, under different network conditions. Our results show that while MPTCP provides useful bandwidth aggregation, congestion avoidance, and improved resiliency for some cloud applications, these benefits do not apply uniformly across applications, especially in cloud settings.

Auto-scalers for clouds promise stable service quality at low costs when facing changing workload intensity. The major public cloud providers provide trigger-based auto-scalers based on thresholds. However, trigger-based auto-scaling has reaction times in the order of minutes. Novel auto-scalers from literature try to overcome the limitations of reactive mechanisms by employing proactive prediction methods. However, the adoption of proactive auto-scalers in production is still very low due to the high risk of relying on a single proactive method. This paper tackles the challenge of reducing this risk by proposing a new hybrid auto-scaling mechanism, called Chameleon, combining multiple different proactive methods coupled with a reactive fallback mechanism. Chameleon employs on-demand, automated time series-based forecasting methods to predict the arriving load intensity in combination with run-time service demand estimation to calculate the required resource consumption per work unit without the need for application instrumentation. We benchmark Chameleon against five different state-of-the-art proactive and reactive auto-scalers one in three different private and public cloud environments. We generate five different representative workloads each taken from different real-world system traces. Overall, Chameleon achieves the best scaling behavior based on user and elasticity performance metrics, analyzing the results from 400 hours aggregated experiment time.

Simplifying the task of resource management and scheduling for customers, while still delivering complex Quality-of-Service (QoS), is key to cloud computing. Many autoscaling policies have been proposed in the past decade to decide on behalf of cloud customers when and how to provision resources to a cloud application utilizing cloud elasticity features. However, in prior work, when a new policy is proposed, it is seldom compared to the state-of-the-art, and is often compared only to static provisioning using a predefined QoS target. This reduces the ability of cloud customers and of cloud operators to choose and deploy an autoscaling policy. In our work, we conduct an experimental performance evaluation of autoscaling policies, using as application model workflows, a commonly used formalism for automating resource management for applications with well-defined yet complex structure. We present a detailed comparative study of general state-of-the-art autoscaling policies, along with two new workflow-specific policies. To understand the performance differences between the 7 policies, we conduct various forms of pairwise and group comparisons. We report both individual and aggregated metrics. Our results highlight the trade-offs between the suggested policies, and thus enable a better understanding of the current state-of-the-art.

Given the complexity and heterogeneity in Cloud computing scenarios, the modeling approach has widely been employed to investigate and analyze the energy consumption of Cloud applications, by abstracting real-world objects and processes that are difficult to observe or understand directly. It is clear that the abstraction sacrifices, and usually does not need, the complete reflection of the reality to be modeled. Consequently, current energy consumption models vary in terms of purposes, assumptions, application characteristics and environmental conditions, with possible overlaps between different research works. Therefore, it would be necessary and valuable to reveal the state-of-the-art of the existing modeling efforts, so as to weave different models together to facilitate comprehending and further investigating application energy consumption in the Cloud domain. By systematically selecting, assessing, and synthesizing 76 relevant studies, we rationalized and organized over 30 energy consumption models with unified notations. To help investigate the existing models and facilitate future modeling work, we deconstructed the runtime execution and deployment environment of Cloud applications, and identified 18 environmental factors and 12 workload factors that would be influential on the energy consumption. In particular, there are complicated trade-offs and even debates when dealing with the combinational impacts of multiple factors.

Rapid elasticity is one of the essential characteristics of cloud computing identified by NIST [17]. Elasticity allows resources to be provisioned and released to scale rapidly out ward and in ward according to demand. Tens -- if not hundreds -- of algorithms have been proposed in the literature to automatically achieve elastic provisioning [15, 23, 14, 21, 13, 20, 6, 12, 16, 10]. These algorithms are typically referred to as elasticity algorithms, dynamic provisioning techniques or autoscalers. While trying to solve the same problem, sometimes with differing assumption, many of these algorithms are either compared to static provisioning or to a predefined QoS target, e.g., predefined response time target, with very little -- or no -- comparison to previously published work. This reduces the ability of an application owner or a cloud operator to choose and deploy a suitable algorithm from the literature. Many of these algorithms have been tested with one single -- real or synthetic -- workload in a specific use-case [13, 14, 10]. While all published algorithms are shown to work in the specific use-case they were designed for with the, typically short, workloads tested with, it is seldom the case that the real scenarios will be any thing close to the test cases for which the algorithms are shown to work. Bursts occur in workloads occasionally. Workload dynamics change over time and the load-mix of an application significantly affects how provisioning should be done [21].

Video-on-Demand (VoD) and video sharing services account for a large percentage of the total downstream Internet traffic. In order to provide a better understanding of the load on these services, we analyze and model a workload trace from a VoD service provided by a major Swedish TV broadcaster. The trace contains over half a million requests generated by more than 20000 unique users. Among other things, we study the request arrival rate, the inter-arrival time, the spikes in the workload, the video popularity distribution, the streaming bit-rate distribution and the video duration distribution. Our results show that the user and the session arrival rates for the TV4 workload does not follow a Poisson process. The arrival rate distribution is modeled using a lognormal distribution while the inter-arrival time distribution is modeled using a stretched exponential distribution. We observe the "impatient user" behavior where users abandon streaming sessions after minutes or even seconds of starting them. Both very popular videos and non-popular videos are particularly affected by impatient users. We investigate if this behavior is an invariant for VoD workloads.

Recent advances in hardware development coupled with the rapid adoption and broad applicability of cloud computing have introduced widespread heterogeneity in data centers, significantly complicating the management of cloud applications and data center resources. This paper presents the CACTOS approach to cloud infrastructure automation and optimization, which addresses heterogeneity through a combination of in-depth analysis of application behavior with insights from commercial cloud providers. The aim of the approach is threefold: to model applications and data center resources, to simulate applications and resources for planning and operation, and to optimize application deployment and resource use in an autonomic manner. The approach is based on case studies from the areas of business analytics, enterprise applications, and scientific computing.

Accurate understanding of workloads is key to efficient cloud resource management as well as to the design of large-scale applications. We analyze and model the workload of Wikipedia, one of the world's largest web sites. With descriptive statistics, time-series analysis, and polynomial splines, we study the trend and seasonality of the workload, its evolution over the years, and also investigate patterns in page popularity. Our results indicate that the workload is highly predictable with a strong seasonality. Our short term prediction algorithm is able to predict the workload with a Mean Absolute Percentage Error of around 2%.

Cloud elasticity is the ability of the cloud infrastructure to rapidly change the amount of resources allocated to a service in order to meet the actual varying demands on the service while enforcing SLAs. In this paper, we focus on horizontal elasticity, the ability of the infrastructure to add or remove virtual machines allocated to a service deployed in the cloud. We model a cloud service using queuing theory. Using that model we build two adaptive proactive controllers that estimate the future load on a service. We explore the different possible scenarios for deploying a proactive elasticity controller coupled with a reactive elasticity controller in the cloud. Using simulation with workload traces from the FIFA world-cup web servers, we show that a hybrid controller that incorporates a reactive controller for scale up coupled with our proactive controllers for scale down decisions reduces SLA violations by a factor of 2 to 10 compared to a regression based controller or a completely reactive controller.