We predict the conditional distributions of service metrics, such as response time or frame rate, from infrastructure measurements in a networked environment. From such distributions, key statistics of the service metrics, including mean, variance, or quantiles can be computed, which are essential for predicting SLA conformance and enabling service assurance. We present and assess two methods for prediction: (1) mixture models with Gaussian or Lognormal kernels, whose parameters are estimated using mixture density networks, a class of neural networks, and (2) histogram models, which require the target space to be discretized. We apply these methods to a VoD service and a KV store service running on our lab testbed. A comparative evaluation shows the relative effectiveness of the methods when applied to operational data. We find that both methods allow for accurate prediction. While mixture models provide a general and elegant solution, they incur a very high overhead related to hyper-parameter search and neural network training. Histogram models, on the other hand, allow for efficient training, but require adjustment to the specific use case.

In networked systems engineering, operational data gathered from sensors or logs can be used to build data-driven functions for performance prediction, anomaly detection, and other operational tasks. The number of data sources used for this purpose determines the dimensionality of the feature space for learning and can reach millions for medium-sized systems. Learning on a space with high dimensionality generally incurs high communication and computational costs for the learning process. In this work, we apply and compare a range of methods, including, feature selection, Principle Component Analysis (PCA), and autoencoders with the objective to reduce the dimensionality of the feature space while maintaining the prediction accuracy when compared with learning on the full space. We conduct the study using traces gathered from a test-bed at KTH that runs a video-on-demand service and a key-value store under dynamic load. Our results suggest the feasibility of reducing the dimensionality of the feature space of operational data significantly, by one to two orders of magnitude in our scenarios, while maintaining prediction accuracy. The findings confirm the Manifold Hypothesis in machine learning, which states that real-world data sets tend to occupy a small subspace of the full feature space. In addition, we investigate the tradeoff between prediction accuracy and prediction overhead, which is crucial for applying the results to operational systems.

We present a framework for achieving end-to-end management objectives for multiple services that concurrently execute on a service mesh. We apply reinforcement learning (RL) techniques to train an agent that periodically performs control actions to reallocate resources. We develop and evaluate the framework using a laboratory testbed where we run information and computing services on a service mesh, supported by the Istio and Kubernetes platforms. We investigate different management objectives that include end-to-end delay bounds on service requests, throughput objectives, cost-related objectives, and service differentiation. Our framework supports the design of a control agent for a given management objective. It is novel in that it advocates a top-down approach whereby the management objective is defined first and then mapped onto the available control actions. Several types of control actions can be executed simultaneously, which allows for efficient resource utilization. Second, the framework separates learning of the system model and the operating region from learning of the control policy. By first learning the system model and the operating region from testbed traces, we can instantiate a simulator and train the agent for different management objectives in parallel. Third, the use of a simulator shortens the training time by orders of magnitude compared with training the agent on the testbed. We evaluate the learned policies on the testbed and show the effectiveness of our approach in several scenarios. In one scenario, we design a controller that achieves the management objectives with $50\%$ less system resources than Kubernetes HPA autoscaling.

Dynamic resource allocation in networked systems is needed to continuously achieve end-to-end management objectives. Recent research has shown that reinforcement learning can achieve near-optimal resource allocation policies for realistic system configurations. However, most current solutions require expensive retraining when changes in the system occur. We address this problem and introduce an efficient method to adapt a given base policy to system changes, e.g., to a change in the service offering. In our approach, we adapt a base control policy using a rollout mechanism, which transforms the base policy into an improved rollout policy. We perform extensive evaluations on a testbed where we run applications on a service mesh based on the Istio and Kubernetes platforms. The experiments provide insights into the performance of different rollout algorithms. We find that our approach produces policies that are equally effective as those obtained by offline retraining. On our testbed, effective policy adaptation takes seconds when using rollout, compared to minutes or hours when using retraining. Our work demonstrates that rollout, which has been applied successfully in other domains, is an effective approach for policy adaptation in networked systems.