Model driven telemetry (MDT) enables the real-time collection of hundreds of thousands of counters on large-scale networks, with contextual information to each counter provided in the telemetry data structure definition. Explaining network events in such datasets implies substantial analysis by a domain expert. This paper presents an semantic feature selection method, to find the most important counters which describe a given event in a telemetry dataset, and facilitate the explanation process. This paper proposes a metric for estimating the importance of features in a dataset with descriptive feature names, to find those that are most meaningful to a human. With this estimation, this paper presents a cross-entropy based metric describing the quality of a selection of counters, which is combined with the data behavior to define an optimization goal. The computation of optimal selections distills intelligible and precise selections of counters with adjustable verbosity, and describes events with a few selected counters outlining the root cause of network events.

Deep neural network (DNN) inference on streaming data requires computing resources to satisfy inference throughput requirements. However, latency and privacy sensitive deep learning applications cannot afford to offload computation to remote clouds because of the implied transmission cost and lack of trust in third-party cloud providers. Among solutions to increase performance while keeping computation on a constrained environment, hardware acceleration can be onerous, and model optimization requires extensive design efforts while hindering accuracy. DNN partitioning is a third complementary approach, and consists of distributing the inference workload over several available edge devices, taking into account the edge network properties and the DNN structure, with the objective of maximizing the inference throughput (number of inferences per second). This paper introduces a method to predict inference and transmission latencies for multi-threaded distributed DNN deployments, and defines an optimization process to maximize the inference throughput. A branch and bound solver is then presented and analyzed to quantify the achieved performance and complexity. This analysis has led to the definition of the acceleration region, which describes deterministic conditions on the DNN and network properties under which DNN partitioning is beneficial. Finally, experimental results confirm the simulations and show inference throughput improvements in sample edge deployments.

No abstract

Expert systems for fault diagnosis are computationally expensive to build and maintain, and lack scalability and inherent adaptability to unknown events or modifications in the topology of the monitored system. While data-driven feature selection mechanisms can facilitate diagnosis without the hardship of developing and maintaining expert systems, purely data-driven mechanisms lack understanding of semantic importance within a feature set, and would benefit from additional domain knowledge. Part of this additional knowledge can be extracted from metadata. The proposed approach combines data-driven metrics and semantic information contained in the feature names to produce selections of features which best represent an underlying event. This study extends a cross entropy based optimization method to join semantic importance with data behavior. A benchmarking architecture is introduced to evaluate the benefits of semantic analysis, and demonstrate the performance and robustness of semantic feature selection on different types of faults in network telemetry datasets, modeled with the YANG data modeling language. The results illustrate the interest of such a complementary meta-data analysis for data-driven fault diagnosis, and highlight the robustness of the studied approach against variations in the input feature set.