A Proposal of an Efficient Path Selection Method Using INT-Based Delay Measurement

A Proposal of an Efficient Path Selection Method Using INT-Based Delay Measurement 1st Kota NAKAGAWA Nara Institute of Science and Technology Ikoma-shi, Nara, Japan nakaga[email protected] 2nd Kohei ICHIKAWA Nara Institute of Science and Technology Ikoma-shi, Nara, Japan ichikaw[email protected] 3rd Hajimu IIDA Nara Institute of Science and Technology Ikoma-shi, Nara, Japan [email protected] Abstract—In this study, we propose an optimal path selection method for source routing based on delay measurement using In-band Network Telemetry (INT). Conventional path selection methods often determine routes based on the minimum number of hops. However, this approach does not necessarily guarantee the selection of the lowest-latency path, as it does not take into account the bandwidth or congestion conditions of each link. In our proposed method, INT packets are duplicated at intermediate nodes and forwarded across multiple paths, enabling efficient delay measurement and selection of the path with the lowest latency. To evaluate the effectiveness of our method, we conducted simulations using the Internet2 OS3E network topology. Compared to the conventional OSPFv3 approach, our method successfully selected lower-latency paths for 73 out of 519 sourcedestination node pairs, excluding directly connected nodes where no alternative path can achieve lower latency. Index Terms—SDN, SRv6, INT, P4 I. INTRODUCTION With the widespread adoption of the Internet, various applications, such as video communication and file transfer, have become increasingly utilized. The performance of these applications is significantly affected by network communication quality. However, conventional routing protocols such as Border Gateway Protocol (BGP) [1] and Open Shortest Path First (OSPF) [2] primarily select the shortest path between the source and destination, which may not always be optimal for application performance. Since conventional routing protocols do not consider application requirements, Software Defined Networking (SDN) has emerged as a solution to enable more fine-grained network management and control. SDN allows for flexible network configuration and management through software, enabling application-aware path selection. In particular, OpenFlow is one of the representative technologies of SDN. Previous studies have proposed methods for selecting optimal paths for applications using OpenFlow [3], demonstrating its effectiveness. OpenFlow enables flexible path control by managing the This work was partially supported by JSPS KAKENHI Grant Numbers 21K11913. flow tables of switches collectively through a central controller. However, because the controller manages all flows centrally, the number of flow table entries increases with the number of applications, leading to scalability issues and increased controller load [4]. To overcome this limitation, Segment Routing (SR) has been proposed. SR eliminates the need for a centralized controller, allowing each node to independently control routing, thereby enabling scalable network operations. In SR, each node or link within the network is assigned a Segment Identifier (SID), and packets include a list of these SIDs to specify the routing path. In particular, Segment Routing over IPv6 (SRv6) [5] extends IPv6 to embed routing information directly into packets, avoiding excessive growth of flow tables as seen in OpenFlow. While SRv6 mitigates flow table scalability issues, effective network monitoring is still crucial for selecting optimal paths that meet application requirements. Conventional monitoring methods, such as using Simple Network Management Protocol (SNMP) [6], involve collecting data from nodes and switches via a monitoring server. However, this centralized approach increases the load on the monitoring server as the network scales. To address these challenges, In-band Network Telemetry (INT) [7] has recently gained attention as a network monitoring technology. INT embeds monitoring metadata directly into packets, allowing nodes to update the information during packet forwarding. This enables real-time traffic monitoring without generating additional monitoring traffic. Moreover, since INT enables end-to-end traffic monitoring, the processing load of monitoring information can be distributed across nodes. This study aims to enhance network monitoring and optimize routing control by leveraging INT with SRv6 technology. Specifically, we propose a method that focuses on network delay measurement using INT to achieve realtime network state monitoring and optimal path selection. Through this approach, we aim to realize a flexible and efficient application-aware network control mechanism based SID１ 2. Return each INT packet along the path in traversed using SRv6 1. Embed a timestamp and send the packet SID２ SID３ SID４ SID５ INT Header SID１ Timestamp SRv6 Header SID３ SID２ SID１ INT Header SID３ SID２ SID１ Timestamp SRv6 Header SID４ SID２ SID１ INT Header SID４ SID２ SID１ Timestamp 3. Calculate the lowest-latency path based on the timestamps Fig. 1. Low-Latency Path Selection Process Using the Proposed Method on SRv6. II. RELATED RESEARCH Yan Zheng et al. proposed a telemetry framework called Multipath In-band Network Telemetry (MPINT) [8], specifically designed for monitoring multicast traffic. Direct application of conventional INT to multicast traffic causes redundant telemetry data collection from the same network devices, leading to bandwidth waste and unnecessary data duplication. To address this issue, MPINT modifies multicast packet replication at intermediate switches by stripping previously collected telemetry data and updating only the path information. The experimental results show that MPINT reduces INT bandwidth overhead by 80% and decreases telemetry data uploaded to controllers by 50% compared to conventional INT. Tomoki Sugiura et al. proposed a system called Acar [9] to overcome the limitations of conventional routing protocols such as BGP and OSPF, which do not consider application-specific path requirements. Acar leverages SRv6 for application-aware routing. It monitors network conditions using SNMP and optimizes routing paths for applications based on collected data. Validation in a virtual network environment demonstrated that Acar improves load balancing between links and enhances throughput compared to conventional Equal-Cost Multipath (ECMP) routing. While MPINT improves telemetry efficiency and Acar optimizes path selection using SRv6 and network monitoring, neither fully leverages INT for real-time delay-aware routing. This study addresses this gap by integrating INT with SRv6 for more efficient and accurate path selection. III. PROPOSED METHOD In this study, we propose a method for dynamically selecting the lowest-latency path based on network conditions, specifically targeting applications that require low latency. In the proposed method, rather than sending INT packets individually, intermediate switches duplicate them to acquire more candidate paths dynamically and enhance network state monitoring. This enables parallel collection of multiple path metrics, facilitating the selection of the optimal path for the application. TABLE I SPECIFICATIONS OF THE PHYSICAL MACHINE Item Specification CPU Intel(R) Xeon(R) Silver 4208 CPU @ 2.10GHz 8 cores × 2 Memory 96GB TABLE II SPECIFICATIONS OF THE VIRTUAL MACHINE Item Specification OS Ubuntu 22.04 CPU Intel(R) Xeon(R) Silver 4208 CPU @ 2.10GHz 12 cores Memory 48GB Figure 1 illustrates the process of selecting a low-latency path using the proposed method. The source node (SID1) sends an INT packet containing a timestamp to the destination node (SID5). After receiving multiple duplicated INT packets, SID5 returns them to SID1 along the paths they originally traversed using SRv6. SID1 then compares the timestamps within the returned INT packets with its own reception time to determine the path with the lowest latency. This allows realtime path selection based on latency information. However, simply duplicating and propagating INT packets may result in excessive collection of redundant path information, potentially causing a broadcast storm. To prevent this issue, the proposed method analyzes the path information within INT packets and discards packets when a loop is detected. This prevents unnecessary packet proliferation and reduces network load. To implement the proposed method, we developed a virtual switch on Mininet consisting of two main components: a P4 switch for defining the processing behavior of INT packets using the Programming Protocol-independent Packet Processors (P4) language [10] and a controller for managing packet routing using FRRouting, a software router. Rather than implementing routing functionality independently in P4, we utilized FRRouting, a widely used routing suite. This approach allows support for various routing protocols, including BGP and OSPF, as well as SRv6, enabling flexible and diverse routing control. IV. EVALUATION A. Experimental Environment In this experiment, we conducted simulations using Mininet on a virtual machine. Tables I and II show the specifications of the physical and virtual machines used in the experiment. To evaluate the effectiveness of the proposed method in a practical setting, we used the topology of Open Science, Scholarship, and Services Exchange (OS3E), a networking platform provided by Internet2 [11] for academic research. The experiment was conducted under conditions that simulate a real-world environment. Internet2 OS3E [12] is an operational network connecting major cities in the United States and Canada, including Vancouver, Seattle, Chicago, New York, Fig. 2. Overall topology of Internet2 OS3E and Los Angeles. Figure 2 illustrates the overall topology of Internet2 OS3E. This topology provides real-world operational data, such as the latitude and longitude of each node and the delay time between links. By constructing an environment that mimics this topology, we can perform simulations in a setting close to a real-world network [13]. Utilizing the topology of an actively operated wide-area network allows for a more realistic performance evaluation of the proposed method, enhancing the reliability of its effectiveness. B. Experimental Details For comparison with the proposed method, we use OSPF version 3 (OSPFv3) [14] as the routing protocol. OSPFv3 and SRv6 settings are configured using FRRouting, with the IPv6 address of the loopback interface set as the OSPFv3 router ID and the SRv6 Segment Identifier (SID) assigned based on the OSPF router ID. The measurement method involves sending UDP packets containing a timestamp from each node and measuring the delay to the destination host node. The effectiveness of the proposed method is evaluated by comparing the end-to-end delay with that of OSPFv3-based routing, following the procedure below. In the proposed method, the source node first sends INT packets to the destination node to explore the low-latency paths and configures both nodes to use the lowest-latency route. After the configuration, the source node generates an SRv6 header and sends packets containing timestamps. The delay is measured by calculating the difference between the timestamp of the received packet at the source node and the timestamp at the time of transmission. For OSPFv3-based routing, all link costs are set to the same value, and the OSPFv3 protocol determines the shortest hop-count path in advance. The source node then sends UDP packets (including timestamps) to the destination node along this predetermined path. The delay is measured by computing the difference between the timestamp at the destination node upon reception and the timestamp at the time of transmission. We conducted ten measurements for both methods and calculated the average delay time for comparison. In the actual experiment, the delay times between links provided in the Internet2 OS3E topology information were increased tenfold and used as simulation settings. The Internet2 OS3E platform is maintained by a single organization (Internet2), and the original link delays are extremely low. Since these delays are significantly smaller than the fluctuations occurring in the Mininet simulation environment, it is difficult to observe the effects of delay optimization through path selection. Therefore, the provided delay values were increased to simulate a widearea network operated by multiple organizations. The modified average link delay was set to 26.278 ms. C. Experimental Results In the topology of Internet2 OS3E, there are 519 sourcedestination node pairs, excluding directly connected pairs. In this experiment, we evaluated and compared the routing results and latency for all these pairs using the proposed method and the existing OSPFv3 routing protocol. The evaluation revealed that there were 73 pairs where the routes selected by the proposed method differed from those selected by OSPFv3, resulting in distinct hop counts. Since OSPFv3 always selects the route with the minimum hop count, this discrepancy indicates that the proposed method chose routes with higher hop counts compared to those selected by OSPFv3. Despite the increase in hop counts, the latency of the routes selected by the proposed method was lower than that of the routes selected by OSPFv3 for all 73 pairs. This result suggests that in geographically large and widely distributed networks, there may exist routes that connect source and destination nodes with lower latency, even if they involve a greater number of intermediate nodes. The proposed method demonstrated its capability to discover such latency-optimized routes, highlighting its potential advantage in optimizing network performance in complex topologies. V. CONCLUSION In this study, we proposed an optimal path control method based on latency measurement using INT. While conventional routing selects paths based on the minimum hop count, the proposed method efficiently measures actual latency by duplicating INT packets and determines paths accordingly. Simulation results demonstrated that the proposed method selects lower-latency paths compared to conventional routing protocols. The proposed method is a decentralized path control approach that combines SRv6 and INT. In the future, it is expected to be applied to flexible network control tailored to the characteristics of various applications. REFERENCES [1] Y. Rekhter, S. Hares, and T. Li, “A Border Gateway Protocol 4 (BGP4).” RFC 4271, Jan. 2006. [2] J. Moy, “OSPF Version 2.” RFC 2328, Apr. 1998. [3] P. U-Chupala, K. Ichikawa, H. Iida, N. Kessaraphong, P. Uthayopas, S. Date, H. Abe, H. Yamanaka, and E. Kawai, “Application-oriented bandwidth and latency aware routing with open flow network,” in 2014 IEEE 6th International Conference on Cloud Computing Technology and Science, pp. 775–780, 2014. [4] M. P. Fernandez, “Comparing openflow controller paradigms scalability: Reactive and proactive,” in 2013 IEEE 27th International Conference on Advanced Information Networking and Applications (AINA), pp. 1009– 1016, 2013. [5] C. Filsfils, P. Camarillo, J. Leddy, D. Voyer, S. Matsushima, and Z. Li, “Segment routing over ipv6 (srv6) network programming.” RFC 8986, feb 2021. [6] M. Fedor, M. L. Schoffstall, J. R. Davin, and D. J. D. Case, “Simple Network Management Protocol (SNMP).” RFC 1157, May 1990. [7] T. P. A. W. Group, “In-band network telemetry (int) dataplane specification version 2.1.” https://p4.org/p4-spec/docs/INT v2 1.pdf, 2020. [8] Y. Zheng, T. Pan, Y. Zhang, E. Song, T. Huang, and Y. Liu, “Multipath in-band network telemetry,” in IEEE INFOCOM 2021 - IEEE Conference on Computer Communications Workshops (INFOCOM WKSHPS), pp. 1–2, 2021. [9] T. Sugiura, K. Takahashi, K. Ichikawa, and H. Iida, “Acar: An application-aware network routing system using srv6,” in 2022 IEEE 19th Annual Consumer Communications Networking Conference (CCNC), pp. 751–752, 2022. [10] P. Bosshart, D. Daly, G. Gibb, M. Izzard, N. McKeown, J. Rexford, C. Schlesinger, D. Talayco, A. Vahdat, G. Varghese, et al., “P4: Programming protocol-independent packet processors,” ACM SIGCOMM Computer Communication Review, vol. 44, no. 3, pp. 87–95, 2014. [11] Internet2, “Internet2.” https://internet2.edu/, 2025 (accessed February 20, 2025). [12] Internet2, “Regional research and education networks - internet2.” https://internet2.edu/network/state-and-regional-r-e-networks/, 2025 (accessed February 20, 2025). [13] J. Zhao, H. Qu, J. hong Zhao, Z. Luan, and Y. Guo, “Towards controller placement problem for software-defined network using affinity propagation,” Electronics Letters, vol. 53, pp. 928–929, 2017. [14] D. Ferguson, A. Lindem, and J. Moy, “OSPF for IPv6.” RFC 5340, July 2008.