[Paper Review] SRv6-PM: Performance Monitoring of SRv6 Networks with a Cloud-Native Architecture

Pierpaolo Loreti, Andrea Mayer+10 more|arXiv (Cornell University)|Jul 15, 2020

Software-Defined Networks and 5G33 references4 citations

TL;DR

SRv6-PM proposes a cloud-native, open-source platform for real-time performance monitoring in SRv6 networks, leveraging eBPF-based per-flow packet loss monitoring and a scalable Big Data stack (Kafka, InfluxDB, Grafana) for data ingestion and visualization. It achieves near-real-time loss detection with only 5.4% throughput degradation on egress nodes, outperforming traditional IPset-based methods.

ABSTRACT

Segment Routing over IPv6 (SRv6 in short) is a networking solution for IP backbones and datacenters. The SRv6 standardization, implementation and research are actively progressing and SRv6 has already been adopted in a number of large scale network deployments. Effective Performance Monitoring (PM) solutions for SRv6 networks are strongly needed. The design, implementation and deployment of such PM solutions span the different planes of a networking architecture: Performance Measurements data (packet loss and delay) needs to be measured (in the Data Plane), the monitored nodes needs to be controlled (in the Control Plane), the measured data needs to be collected (in the Control/Management Planes), then the Data must be processed and stored, using Big-Data processing solutions. We focus on Loss Monitoring, by considering a solution capable of tracking single packet loss events in near-real time (e.g. with a delay in the order of 20 seconds). We describe SRv6-PM, a solution for Performance Monitoring of SRv6 networks. SRv6-PM features a cloud-native architecture for the SDN-based control of Linux routers and for ingestion, processing, storage and visualization of PM data. In the Data Plane, SRv6-PM includes efficient building blocks for packet loss evaluation (e.g. the packet counting components) in a Linux router. SRv6-PM is released as open source. Not only we provide a reproducible environment for PM experiments, but we also offer a re-usable and extensible cloud-native platform that can be automatically deployed in different environments, from a single host to multiple servers on private/public clouds.

Motivation & Objective

• To address the lack of scalable, real-time performance monitoring solutions for SRv6 networks, especially for loss monitoring at fine granularity.
• To design and implement a cloud-native architecture that integrates SDN control, data plane measurement, and Big Data processing for end-to-end monitoring.
• To develop an efficient, low-overhead packet loss monitoring mechanism in Linux routers using eBPF, outperforming legacy Netfilter/IPset approaches.
• To provide a reusable, open-source platform for prototyping and standardization of SRv6 PM solutions, supporting future extensions like delay monitoring.

Proposed method

• Extends the TWAMP protocol for SRv6 by using DSCP byte marking to enable single-packet loss detection via alternate marking.
• Implements an eBPF-based Per-Flow Packet Loss Monitoring (PF-PLM) module in the Linux kernel's XDP path for low-latency, high-throughput packet processing.
• Designs a gRPC-based API between the SDN controller and network nodes to configure and retrieve per-flow loss measurements.
• Deploys a cloud-native data pipeline using Kafka for streaming, Telegraf for metrics collection, InfluxDB for time-series storage, and Grafana for real-time visualization.
• Uses ArangoDB for structured storage of monitoring metadata and configuration data, enabling efficient querying and correlation.
• Deploys the entire platform in a containerized, Kubernetes-native environment for portability across on-premise and public cloud infrastructures.

Experimental results

Research questions

• RQ1 How can SRv6 networks achieve near-real-time, per-flow packet loss monitoring with minimal performance overhead in the data plane?
• RQ2 What is the performance impact of eBPF-based packet marking and counting compared to traditional Netfilter/IPset-based solutions in Linux routers?
• RQ3 How can a cloud-native, open-source architecture efficiently scale to collect, process, and visualize performance data from hundreds of SRv6 nodes?
• RQ4 Can the proposed platform accurately detect and report single-packet loss events with high fidelity, matching application-level measurements?
• RQ5 How reusable and extensible is the platform for future PM extensions, such as SRv6 delay monitoring?

Key findings

• The eBPF-based PF-PLM implementation achieves only 5.4% average throughput degradation on the egress node when monitoring one flow, with stable performance regardless of the number of monitored flows.
• The eBPF solution outperforms the IPset-based alternative by reducing degradation from 17.6% to 5.4% on the egress node, demonstrating superior scalability and efficiency.
• The SRv6-PM platform successfully detects and reports single-packet loss events with high accuracy, matching application-level measurements from iPerf with negligible deviation.
• The platform achieves near-real-time monitoring with a delay of approximately 20 seconds, meeting the target for operational network monitoring.
• The open-source, cloud-native design enables automatic deployment across diverse environments, from single hosts to multi-node cloud clusters, ensuring reproducibility and extensibility.
• The integration of Kafka, InfluxDB, and Grafana enables low-latency data ingestion, efficient storage, and real-time visualization of performance metrics.