Joint Minimization of Monitoring Cost and Delay in Overlay Networks: Optimal Policies with a Markovian Approach

Abstract

Continuous monitoring of network resources enables to make more-informed resource allocation decisions but incurs overheads. We investigate the trade-off between monitoring costs and benefits of accurate state information for a routing problem. In our approach link delays are modeled by Markov chains or hidden Markov models. The current delay information on a link can be obtained by actively monitoring this link at a fixed cost. At each time slot, the decision maker chooses to monitor a subset of links with the objective of minimizing a linear combination of long-run average delay and monitoring costs. This decision problem is modeled as a Markov Decision Process whose solution is computed numerically. In addition, in simple setttings we prove that immediate monitoring cost and delay minimization leads to a threshold policy on a filter which sums up information from past measurements. The lightweight method as well as the optimal policy are tested on several use-cases. We demonstrate on an overlay of 30 nodes of RIPE Atlas that we obtain delay values close to the performance of the always best path with an extremely low monitoring effort when delays between nodes are modeled with hierarchical Dirichlet process hidden Markov models.

Scalable Monitoring Heuristics for Improving Network Latency

“…More recently, in [10] we introduce a theoretical framework for taking monitoring and routing decisions that offer a good trade-off between monitoring costs and the gain in delay from appropriate routing decisions. This framework is based on Markov Decision Processes (MDPs) and a markovian modeling of path delays.…”

“…However, as pointed out, the exact computation of the optimal monitoring policy is only feasible for small-scale examples. In the present paper, we adopt the theoretical framework introduced in [10] and propose scalable monitoring methods which can be used for larger scenarios.…”

“…In [10], the trade-off between accurate latency information and monitoring costs was investigated in the framework of an MDP. However, computing the exact solution of this MDP is prohibitively expensive and not practical when the number of paths for each OD pair is not very small.…”

Performance analysis of container-based networking solutions for high-performance computing cloud

“…An overlay network [16] is the most commonly used container network on multiple nodes. An overlay network configuration is needed to allow containers to communicate with each other on multiple nodes [17][18][19]. Representative overlay networking solutions for containers are divided into two catagories: linux bridge overlay network and Flannel networking.…”

Poster Abstract: A flexible infinite HMM model for accurate characterization and segmentation of RTT timeseries

“…We managed to reduce the number of measurements by 90% while having on average an optimized path only 0.65ms longer than the optimal path. We refer the reader to [4] for more information.…”