Enda Barrett

Applying reinforcement learning towards automating resource allocation and application scalability in the cloud

Public Infrastructure as a Service (IaaS) clouds such as Amazon, GoGrid and Rackspace deliver computational resources by means of virtualisation technologies. These technologies allow multiple independent virtual machines to reside in apparent isolation on the same physical host. Dynamically scaling applications running on IaaS clouds can lead to varied and unpredictable results because of the performance interference effects associated with co‐located virtual machines. Determining appropriate scaling policies in a dynamic non‐stationary environment is non‐trivial. One principle advantage exhibited by IaaS clouds over their traditional hosting counterparts is the ability to scale resources on‐demand. However, a problem arises concerning resource allocation as to which resources should be added and removed when the underlying performance of the resource is in a constant state of flux. Decision theoretic frameworks such as Markov Decision Processes are particularly suited to decision making under uncertainty. By applying a temporal difference, reinforcement learning algorithm known as Q‐learning, optimal scaling policies can be determined. Additionally, reinforcement learning techniques typically suffer from curse of dimensionality problems, where the state space grows exponentially with each additional state variable. To address this challenge, we also present a novel parallel Q‐learning approach aimed at reducing the time taken to determine optimal policies whilst learning online. Copyright

A Learning Architecture for Scheduling Workflow Applications in the Cloud

The scheduling of workflow applications involves the mapping of individual workflow tasks to computational resources, based on a range of functional and non-functional quality of service requirements. Workflow applications such as scientific workflows often require extensive computational processing and generate significant amounts of experimental data. The emergence of cloud computing has introduced a utility-type market model, where computational resources of varying capacities can be procured on demand, in a pay-per-use fashion. In workflow based applications dependencies exist amongst tasks which requires the generation of schedules in accordance with defined precedence constraints. These constraints pose a difficult planning problem, where tasks must be scheduled for execution only once all their parent tasks have completed. In general the two most important objectives of workflow schedulers are the minimisation of both cost and make span. The cost of workflow execution consists of both computational costs incurred from processing individual tasks, and data transmission costs. With scientific workflows potentially large amounts of data must be transferred between compute and storage sites. This paper proposes a novel cloud workflow scheduling approach which employs a Markov Decision Process to optimally guide the workflow execution process depending on environmental state. In addition the system employs a genetic algorithm to evolve workflow schedules. The overall architecture is presented, and initial results indicate the potential of this approach for developing viable workflow schedules on the Cloud.

A network aware approach for the scheduling of virtual machine migration during peak loads

Live virtual machine migration can have a major impact on how a cloud system performs, as it consumes significant amounts of network resources such as bandwidth. Migration contributes to an increase in consumption of network resources which leads to longer migration times and ultimately has a detrimental effect on the performance of a cloud computing system. Most industrial approaches use ad-hoc manual policies to migrate virtual machines. In this paper, we propose an autonomous network aware live migration strategy that observes the current demand level of a network and performs appropriate actions based on what it is experiencing. The Artificial Intelligence technique known as Reinforcement Learning acts as a decision support system, enabling an agent to learn optimal scheduling times for live migration while analysing current network traffic demand. We demonstrate that an autonomous agent can learn to utilise available resources when peak loads saturate the cloud network.

A parallel framework for Bayesian reinforcement learning

Solving a finite Markov decision process using techniques from dynamic programming such as value or policy iteration require a complete model of the environmental dynamics. The distribution of rewards, transition probabilities, states and actions all need to be fully observable, discrete and complete. For many problem domains, a complete model containing a full representation of the environmental dynamics may not be readily available. Bayesian reinforcement learning (RL)\ is a technique devised to make better use of the information observed through learning than simply computing Q-functions. However, this approach can often require extensive experience in order to build up an accurate representation of the true values. To address this issue, this paper proposes a method for parallelising a Bayesian RL technique aimed at reducing the time it takes to approximate the missing model. We demonstrate the technique on learning next state transition probabilities without prior knowledge. The approach is general enough for approximating any probabilistically driven component of the model. The solution involves multiple learning agents learning in parallel on the same task. Agents share probability density estimates amongst each other in an effort to speed up convergence to the true values.

A reinforcement learning approach for dynamic selection of virtual machines in cloud data centres

In recent years Machine Learning techniques have proven to reduce energy consumption when applied to cloud computing systems. Reinforcement Learning provides a promising solution for the reduction of energy consumption, while maintaining a high quality of service for customers. We present a novel single agent Reinforcement Learning approach for the selection of virtual machines, creating a new energy efficiency practice for data centres. Our dynamic Reinforcement Learning virtual machine selection policy learns to choose the optimal virtual machine to migrate from an over-utilised host. Our experiment results show that a learning agent has the abilities to reduce energy consumption and decrease the number of migrations when compared to a state-of-the-art approach.

An Autonomous Network Aware VM Migration Strategy in Cloud Data Centres

Live virtual machine migration can have a major impact on how a cloud system performs, as it can consume significant amounts of network resources such as bandwidth. A virtual machine migration occurs when a host becomes over-utilised or under utilised. Migration contributes to an increase in consumption of network resources which leads to longer migration times and ultimately has a detrimental effect on the performance of a cloud system. In this paper, we propose an autonomous network aware virtual machine migration strategy that observes the current demand level of a network and performs appropriate actions based on what it is experiencing. The Artificial Intelligence technique known as Reinforcement Learning acts as a decision support system, enabling an agent to learn an optimal time to schedule a virtual machine migration depending on the current network traffic demand. We show that an autonomous agent can learn to utilise available network resources when network saturation occurs at peak times.

Autonomous HVAC Control, A Reinforcement Learning Approach

Recent high profile developments of autonomous learning thermostats by companies such as Nest Labs and Honeywell have brought to the fore the possibility of ever greater numbers of intelligent devices permeating our homes and working environments into the future. However, the specific learning approaches and methodologies utilised by these devices have never been made public. In fact little information is known as to the specifics of how these devices operate and learn about their environments or the users who use them. This paper proposes a suitable learning architecture for such an intelligent thermostat in the hope that it will benefit further investigation by the research community. Our architecture comprises a number of different learning methods each of which contributes to create a complete autonomous thermostat capable of controlling a HVAC system. A novel state action space formalism is proposed to enable a Reinforcement Learning agent to successfully control the HVAC system by optimising both occupant comfort and energy costs. Our results show that the learning thermostat can achieve cost savings of 10% over a programmable thermostat, whilst maintaining high occupant comfort standards.

A reinforcement learning approach for the scheduling of live migration from under utilised hosts

Live virtual machine migration can have a major impact on how a cloud system performs, as it consumes significant amount of network resources, such as bandwidth. A virtual machine migration occurs when a host becomes over-utilised or under-utilised. In this paper, we propose a network aware live migration strategy that monitors the current demand level of bandwidth when network congestion occurs and performs appropriate actions based on what it is experiencing. The Artificial Intelligence technique that is based on Reinforcement Learning acts as a decision support system, enabling an agent to learn an optimal time to schedule a virtual machine migration depending on the current bandwidth usage in a data centre. We show from our results that an autonomous agent can learn to utilise available network resources such as bandwidth when network saturation occurs at peak times.

Single system image

Single system image is a computing paradigm where a number of distributed computing resources are aggregated and presented via an interface that maintains the illusion of interaction with a single system. This approach encompasses decades of research using a broad variety of techniques at varying levels of abstraction, from custom hardware and distributed hypervisors to specialized operating system kernels and user-level tools. Existing classification schemes for SSI technologies are reviewed, and an updated classification scheme is proposed. A survey of implementation techniques is provided along with relevant examples. Notable deployments are examined and insights gained from hands-on experience are summarized. Issues affecting the adoption of kernel-level SSI are identified and discussed in the context of technology adoption literature. We provide a retrospective survey of single system image.There has been novel recent work in the area of distributed hypervisors.Despite a peak in interest in the 2000s, kernel-level SSI has not been widely adopted.There may be a role in the future for virtualized kernel-level SSI clusters.

Predicting host CPU utilization in the cloud using evolutionary neural networks

Abstract The Infrastructure as a Service (IaaS) platform in cloud computing provides resources as a service from a pool of compute, network, and storage resources. One of the major challenges facing cloud computing is to predict the usage of these resources in real time. By knowing future demands, cloud data centres can dynamically scale resources to decrease energy consumption while maintaining a high quality of service. However cloud resource consumption is ever changing, making it difficult for accurate predictions to be produced. This motivates the research presented in this paper which aims to predict in advance the level of CPU consumption of a host. This research implements evolutionary Neural Networks (NN), a powerful machine learning method, to make these predictions. A number of state of the art swarm and evolutionary optimization algorithms are implemented to train the neural networks to predict host utilization: Particle Swarm Optimization (PSO), Differential Evolution (DE) and Covariance Matrix Adaptation Evolutionary Strategy (CMA-ES). The results of this research demonstrate that CMA-ES converges faster to a better solution on the training data. However when evaluated on the test data, DE performs statistically equal to CMA-ES. The results also demonstrate that the trained networks are still accurate when applied to CPU utilization data from different hosts with no further training needed. When evaluated to predict multiple steps into the future, the accuracy of the network understandably decreases but still performs well on average.