Probe-Optimizer: Discovering important nodes for proactive in-band network telemetry to achieve better probe orchestration

Abstract

By embedding the state data maintained by the programmable data plane into additional customizable probes, proactive in-band network telemetry (INT) can easily achieve flexible, full-coverage and fine-grained network measurement. However, a significant portion of these probes are invalid, failing to capture meaningful network event information, and instead increasing bandwidth occupancy as well as communication overhead between the control plane and the data plane. Furthermore, these invalid probes exacerbate controller overhead, forcing resource-limited CPUs to perform a large amount of meaningless computation and analysis. In this paper, we propose Probe-Optimizer, a novel framework tailored for proactive INT, which can reduce the introduction of invalid probes to comprehensively lower the various telemetry overheads mentioned above. Technically, Probe-Optimizer assigns a unique importance to each node in the telemetry scenario. The importance is significantly related to the probability of network events occurring, which can be used to select important nodes worth monitoring in the topology over a period of time. Then, Probe-Optimizer generates a dedicated set of probe paths for important nodes and another set for the remaining nodes/links, customizing a more appropriate probe frequency for each probe path. Extensive evaluations on both random and FatTree topologies with different scales are conducted. The results show that Probe-Optimizer introduces significantly fewer invalid probes. Benefiting from this, for the topology with a size of more than 200 nodes, compared to the state-of-art proactive INT methods, Probe-Optimizer achieves a higher proportion of probes carrying network events and at least 13%, 42%, and 26% lower communication overhead, CPU usage, and average bandwidth occupancy, respectively.