Operating large scale and feature-rich applications is becoming increasingly complex as engineers need to deploy highly configurable software releases on distributed cloud stacks while managing ever-shorter production cycles. Although recent proposals attempt to streamline cloud resources orchestration, there is still a significant challenge in making such solutions generalize to unseen cloud stacks. In other words, the behavior of application-specific Key Performance Indicators (KPIs) and resource configurations, crafted for specific stacks, may differ on heterogeneous deployments, requiring time-consuming policy adjustments. We introduce IMITA, a system that leverages imitation learning to create models by imitating an expert behavior that can be generalized seamlessly to new cloud stacks. To make a generalized model, IMITA maps expert actions taken based on the application KPI space to the space of resource utilization metrics that are universally available in cloud platforms. This mapping enables the model to trigger actions, mimicking expert behavior, upon the occurrence of similar re-source utilization footprints across deployments. We demonstrate IMITA by learning to scale-out Cassandra deployments with diverse configurations and workloads. Our results show IMITA can replicate expert actions across deployments and extrapolate to unseen environments by achieving 50 - 94% fewer false positives actions than traditional threshold-based policies while still adhering to Service-Level Objectives (SLO) and avoiding under-provisioning of resources. Moreover, since collecting data in clouds is costly, IMITA gathers data only for representative configurations to train the imitator model. This approach reduces the size of the collected data to 50%.

Time division multiple access (TDMA) is the medium access control strategy of choice for multihop networks with deterministic delay guarantee requirements. As such, many Internet of Things applications use protocols based on time division multiple access. Optimal slot assignment in such networks is NP-hard when there are strict deadline requirements and is generally done using heuristics that give suboptimal transmission schedules in linear time. However, existing heuristics make a scheduling decision at each time slot based on the same criterion without considering its effect on subsequent network states or scheduling actions. Here, we first identify a set of node features that capture the information necessary for network state representation to aid building schedules using Reinforcement Learning (RL). We then propose three different centralized approaches to RL-based TDMA scheduling that vary in training and network representation methods. Using RL allows applying diverse criteria at different time slots while considering the effect of a scheduling action on meeting the scheduling objective for the entire TDMA frame, resulting in better schedules. We compare the three proposed schemes in terms of how well they meet the scheduling objectives and their applicability to networks with memory and time constraints. One of the schemes proposed is RLSchedule, which is particularly suited to constrained networks. Simulation results for a variety of network scenarios show that RLSchedule reduces the percentage of packets missing deadlines by up to 60% compared to the best available baseline heuristic.

The energy consumption of mobile networks is already substantial nowadays, and only expected to further increase with the roll-out of 5G. Base stations are the key elements in this context: reducing their energy consumption is of paramount importance for network operators, not only to lower operating costs, but also to meet sustainable development goals. Today's base stations are typically over-provisioned, i.e., they comprise multiple cells to meet the peak load in a region. Therefore, substantial energy savings are possible by switching off cells that are under-utilized. This article proposes a data-driven approach to determine the time periods when a cell can be switched off. Forecasting is used to accurately predict network utilization and automatically find the time intervals to reliably switch off a cell. We carefully analyze the requirements of the system as a whole, from data collection to forecasting methods, to enable effective energy savings in practice. Considering several real-world traces from LTE networks, we show that an average of 10.24% energy savings is possible. We explore the trade-offs between energy savings and overhead in switching off cells, and provide insights into the choice of methods accordingly. In particular, we show that the accuracy of forecasting is not the most important factor in achieving energy savings; instead, the prediction (uncertainty) interval plays a key role in being able to achieve energy savings with less impact on end-users. Finally, we propose a model to generate utilization traces that match the distribution of real-world traces obtained from cellular networks.

The adoption of fast-paced practices for developing virtual network functions (VNFs) allows for continuous software delivery and creates a market advantage for network operators. This adoption, however, is problematic for testing engineers that need to assure, in shorter development cycles, certain quality of highly-configurable product releases running on heterogeneous clouds. Machine learning (ML) can accelerate testing workflows by detecting performance issues in new software builds. However, the overhead of maintaining several models for all combinations of build types, network configurations, and other stack parameters, can quickly become prohibitive and make the application of ML infeasible.

The virtual resources of 5G networks are expected to scale and support migration to other locations within the substrate. In this context, a configuration for 5G network slices details the instantaneous mapping of the virtual resources across all slices on the substrate, and a feasible configuration satisfies the Service-Level Objectives (SLOs) without overloading the substrate. Reconfiguring a network from a given source configuration to the desired target configuration involves identifying an ordered sequence of feasible configurations from the source to the target. The proposed solutions for finding such a sequence are optimized for data centers and cannot be used as-is for reconfiguring 5G network slices. We present Matryoshka, our divide-and-conquer approach for finding a sequence of feasible configurations that can be used to reconfigure 5G network slices. Unlike previous approaches, Matryoshka also considers the bandwidth and latency constraints between the network functions of network slices. Evaluating Matryoshka required a dataset of pairs of source and target configurations. Because such a dataset is currently unavailable, we analyze proof of concept roll-outs, trends in standardization bodies, and research sources to compile an input dataset. On using Matryoshka on our dataset, we observe that it yields close-to-optimal reconfiguration sequences 10X faster than existing approaches.

As Multi-access Edge Computing (MEC) and 5G technologies evolve, new applications are emerging with unprecedented capacity and real-time requirements. At the core of such applications there is a need for machine learning (ML) to create value from the data at the edge. Current ML systems transfer data from geo-distributed streams to a central datacenter for modeling. The model is then moved to the edge and used for inference or classification. These systems can be ineffective because they introduce significant demand for data movement and model transfer in the critical path of learning. Furthermore, a full model may not be needed at each edge location. An alternative is to train and update the models online at each edge with local data, in isolation from other edges. Still, this approach can worsen the accuracy of models due to reduced data availability, especially in the presence of local data shifts.

Today, software operation engineers rely on application key performance indicators (KPIs) for sizing and orchestrating cloud resources dynamically. KPIs are monitored to assess the achievable performance and to configure various cloud-specific parameters such as flavors of instances and autoscaling rules, among others. Usually, keeping KPIs within acceptable levels requires application expertise which is expensive and can slow down the continuous delivery of software. Expertise is required because KPIs are normally based on application-specific quality-of-service metrics, like service response time and processing rate, instead of generic platform metrics, like those typical across various environments (e.g., CPU and memory utilization, I/O rate, etc.)

Cloud computing is continuously increasing its popularity as key features such as scalability, pay-per-use and availability continue to evolve. It is also becoming a competitive platform for running high performance computing (HPC) and parallel applications due to the increasing performance of virtualized, highly-available instances. However, migrating HPC applications to cloud still requires native fault-tolerant solutions to fully leverage cloud features and maximize the resource utilization at the best cost - particularly for long-running parallel applications where faults can cause invalid states or data loss. This requires re-executing applications which increases completion time and cost. We propose Resilience as a Service (RaaS), a fault tolerant framework for HPC applications running in cloud. In this paper RADIC architecture (Redundant Array of Distributed Independent Fault Tolerance Controllers) is used to provide clouds with a highly available, distributed and scalable fault-tolerant service. The paper explores how traditional HPC protection and recovery mechanisms must be redesigned to natively leverage cloud properties and its multiple alternatives for implementing rollback recovery protocols using virtual machines, containers, object and block storage or database services. Results show that RaaS restores and completes the application execution using available resources while reducing overhead up to 8% for different fault-tolerant configuration alternatives.

No abstract available.

The virtualization of services with high-availability requirements calls to revisit traditional operation and provisioning processes. Providers are realizing services in software on virtual machines instead of using dedicated appliances to dynamically adjust service capacity to changing demands. Cloud orchestration systems control the number of service instances deployed to make sure each service has enough capacity to meet incoming workloads. However, determining the suitable build-out of a service is challenging as it takes time to install new instances and excessive re-configurations (i.e. scale in/out) can lead to decreased stability. In this paper we present AidOps, a cloud orchestration system that leverages machine learning and domain-specific knowledge to predict the traffic demand, optimizing service performance and cost. AidOps does not require a conservative provisioning of services to cover for the worst-case demand and significantly reduces operational costs while still fulfilling service quality expectations. We have evaluated our framework with real traffic using an enterprise application and a communication service in a private cloud. Our results show up to 4X improvement in service performance indicators compared to existing orchestration systems. AidOps achieves up to 99.985% availability levels while reducing operational costs at least by 20%.

Onboarding network functions onto current clouds requires labor-intensive configuration of the virtual environment. Developers need to dimension the resources available to each virtual machine such as CPU and memory, define thresholds for scaling dynamically and create configuration files that operators can use to execute the network services. This process is time consuming and dependent on the server architecture. As resources are managed on an individual virtual machine basis, services cannot be orchestrated end to end without significant expertise. In this paper, we argue that much of the manual configuration needed for onboarding services onto a cloud can be automated. Moreover, we can automatically generate abstractions that consider services end-to-end and enable their holistic orchestration. We propose a framework that benchmarks network services during the onboarding process and generates an elastic model which relates workload mixes to resource requirements, identifies component dependencies and automates service operation on heterogeneous stacks. We have evaluated our framework using a real-time communication service that handles multiple classes of workloads. Results show that underprovisioning can be eliminated for regular daily traffic, reducing resource provisioning time by at least 5X for the most stressing traffic surges, while improving key performance indicators by at least 40%.

Contemporary Cloud Computing infrastructures are being challenged by an increasing demand for evolved cloud services characterised by heterogeneous performance requirements including real-time, data-intensive and highly dynamic workloads. The classical way to deal with dynamicity is to scale computing and network resources horizontally. However, these techniques must be coupled effectively with advanced routing and switching in a multi-path environment, mixed with a high degree of flexibility to support dynamic adaptation and live-migration of virtual machines (VMs). We propose a management strategy to jointly optimise computing and networking resources in cloud infrastructures, where Software Defined Networking (SDN) plays a key enabling role.

In this paper, we present a brokering logic for providing precise end-to-end QoS levels to cloud applications distributed across a number of different business actors, such as network service providers (NSP) and cloud providers (CSP). The broker composes a number of available offerings from each provider, in a way that respects the QoS application constraints while minimizing costs incurred by cloud consumers. Copyright © 2014 SCITEPRESS - Science and Technology Publications.

The recent trend for Software as a Service and other types of cloud services is driving demand for data centers of ever increasing scale. This will require the scaling up and scaling out of existing data center architectures and will ultimately need new architectures featuring new topologies for the data center networks that interconnect its constituent nodes. The preferred tool for the detailed evaluation of such designs is simulation, as it gives finer detail than analysis at lower cost than test bed evaluation. Realistic simulation times require that the simulation itself be capable of being scaled out, i.e., it should be amenable to parallelization. We describe a simulation framework, together with a methodology for partitioning the relevant simulation models to allow their parallel implementation, and demonstrate its validity by applying it to the simulation of a hybrid optical/electrical network architecture using a cluster of high-end servers. We report on results capturing the performance of our simulator and discuss how these are associated with the underlying hardware hosting the simulation.

Current parallel applications in parallel computing systems require an interconnection network to provide low and bounded communication delays. Communication characteristics such as traffic pattern and communication load change over time and, eventually, they may exceed available network capacity causing congestion and performance degradation. Congestion control based on adaptive routing should be applied in order to adapt quickly to changing traffic conditions. Studies on a vast range of parallel applications show repetitive behavior and can be characterized by a set of representative phases. This work presents a Predictive and Distributed Routing Balancing technique (PR-DRB) to control network congestion based on adaptive traffic distribution. PR-DRB uses speculative routing based on application repetitiveness. PR-DRB monitors messages latencies on routers and logs solutions to congestion, to quickly respond in future similar situations. Experimental results show that the predictive approach could be used to improve performance.

In high performance clusters current parallel application communication needs such as traffic pattern, communication volume, etc., change along time and are difficult to know in advance. Such needs often exceed or do not match available resources causing resource use imbalance, network congestion, throughput reduction and message latency increase, thus degrading the overall system performance. Studies on parallel applications show repetitive behavior that can be characterized by a set of representative phases. This work presents a Predictive and Distributed Routing Balancing (PRDRB) technique, a new method developed to gradually control network congestion, based on paths expansion, traffic distribution, applications pattern repetitiveness and speculative adaptive routing, in order to maintain low latency values. PRDRB monitors messages latencies on routers and logs solutions to congestion, to quickly respond in future similar situations. Traffic congestion experiments were conducted in order to evaluate the performance of the method, and improvements were observed.

The intensive and continuous use of high-performance computing systems for executing computationally intensive applications, coupled with the large number of elements that make them up, dramatically increase the likelihood of failures during their operation. The interconnection network is a critical part of such systems, therefore, network faults have an extremely high impact because most routing algorithms are not designed to tolerate faults. In such algorithms, just a single fault may stall messages in the network, preventing the finalization of applications, or may lead to deadlocked configurations. This paper introduces a novel fault-tolerant routing method provided with a new deadlock avoidance technique designed to solve an unbounded number of faults appearing at random during system operation. Our method provides escape paths for the stalled messages. In addition, the routing algorithm configures alternative paths to avoid the faulty areas taking advantage of communication path redundancy by means of multipath routing approaches. Deadlock avoidance is achieved by adding a small-sized queue and applying a simple set of actions when accessing output buffers with limited free space. Experiments show that our method allows applications to successfully finalize their execution in the presence of several number of faults, with an average performance value of 96% compared to the fault-free scenarios.

The intensive and continuous use of high-performance computing systems for executing computationally intensive applications, coupled with the large number of elements that make them up, dramatically increase the likelihood of failures during their operation. Clearly, network faults have an extremely high impact because most routing algorithms are not designed to tolerate faults. In such algorithms, just a single fault may lead to deadlocked configurations thus preventing the correct finalization of applications. This paper introduces a new deadlock avoidance mechanism for routing algorithms designed to deal with multiple dynamic faults. The mechanism is based on adding a small-sized buffer and applying a simple set of actions when accessing output buffers with limited free space. Unlike typical static solutions, this proposal allows the design of routing algorithms capable of treating an unbounded number of dynamic faults.

The increasing demand of parallel applications in cluster computing requires the use of interconnection networks to provide low and bounded communication delays. However, message congestion appears when communication load between nodes is not fairly distributed over the network. Congestion spreading increases latency and reduces network throughput causing important performance degradation. In this paper we present dynamic routing balancing with multipath distribution (DRB-MD), a new method developed to control network congestion based on a uniform balancing of communication load. DRB-MD distributes the traffic load according to a gradual and load-controlled path expansion. It monitors message latency in network switches, makes decisions about how many alternative paths should be used, and finally decides which path (or paths) to use between each source-destination pair. Experiments with permutation patterns and hotspot traffic were conducted to evaluate DRB-MD performance under conditions commonly created by parallel scientific applications.

Communication requirements in High Performance Computing systems demand the use of high-speed Interconnection networks to connect processing nodes. However, when communication load is unfairly distributed across the network resources, message congestion appears. Congestion spreading increases latency and reduces network throughput causing important performance degradation. The Fast-Response Dynamic Routing Balancing (FR-DRB) is a method developed to perform a uniform balancing of communication load over the interconnection network. FR-DRB distributes the message traffic based on a gradual and load-controlled path expansion. The method monitors network message latency and makes decisions about the number of alternative paths to be used between each source-destination pair for message delivery. FR-DRB performance has been compared with other routing policies under a representative set of traffic patterns which are commonly created by parallel scientific applications. Experiments results show an important improvement in latency and throughput.

Modeling Interconnection networks is an important research topic enabling the study of the interconnection behavior and its significance in telecommunication applications and distributed systems. However, complexity of large-scale networks makes development of models and simulation tools a prohibitively difficult task. In this paper we have explored the network modeling space design to provide models following two different approaches: accurate simulation models based on finite state machines (FSM), and also, analytical models to provide profitable speedup with a minimal accuracy loss. Experiments results show that the proposed analytical model provides a faithful abstraction for the scale of systems that are of interest in the foreseeable future, it reaches an 8% error and speedup of around 30× vs. a FSM model.