Low-Overhead Packet Loss Diagnosis for Virtual Private Clouds using P4-Programmable NICs

Journal Title
Journal ISSN
Volume Title
The virtual private cloud, a logical division into tenants of a service provider's multi-tenant multi-cloud architecture, has become a major trend in the cloud computing industry because of its cost-effectiveness. Effective monitoring systems are critical for cloud tenants to identify and troubleshoot problems that cause negative impact on the reliability and availability of their services in the virtualized cloud infrastructure. However, the virtualized and complex nature of cloud infrastructure limits tenants' ability to pinpoint problems that result in performance degradation, such as packet loss, within the cloud environment. Current monitoring systems are either designed for the physical network or are not adequate to be able to detect the specific cause of packet losses in the physical network. Besides, there are few virtual network monitoring systems suitable for tenants that impose an insignificant overhead. This thesis presents a Packet Loss Diagnosis (PLD) system, which is a specific monitoring service designed to detect packet losses on virtual networks. It notifies virtual private cloud tenants with diagnostic information for troubleshooting. This comprehensive mechanism is based on the modern capabilities of P4 data plane programmable NICs and has a limited footprint in the network. Virtual networks at cloud-scale pose considerable challenges in achieving this goal. We differentiate between packet losses in virtual and physical networks to address these challenges. As for the former, we support tenants with comprehensive information that allows them to locate and resolve their problems on their own. In relation to the latter, we provide high-level information respecting service isolation and abstraction. Furthermore, PLD is designed to meet the requirements of a monitoring system designed for multi-tenant clouds at large scales. To prove the viability of the proposed scheme, we implemented it in P4 as well as investigated its performance through extensive experiments and Mininet simulation. The results of our study demonstrate that the proposed scheme can detect all packet losses reliably and can notify tenants in real-time of the occurrence of packet losses while imposing a minimal overhead that is proportional to the number of packet losses.
Computer Networks, Virtual networks, Datacenter Networks, Programmable Dataplane, Monitoring Systems
Aali Bagi, S. (2023). Low-overhead Packet Loss Diagnosis for virtual private clouds using P4-programmable NICs (Master's thesis, University of Calgary, Calgary, Canada). Retrieved from https://prism.ucalgary.ca.