Master thesis : Intelligent Load Balancing with IOAM and SRv6

Abstract

Load balancing in network and data-centres consists in distributing the traffic as evenly as possible across links and servers. This works aim at proving that a new telemetry tool, IOAM, can be used to measure the current status of a network and use the information collected together with segment routing to implement a dynamic load balancing. A load balancing program that relies on IOAM measurement of a network node queue length was successfully developed and tested to demonstrate feasibility and identify key attention points.

To test the program, two virtual topologies were implemented using linux namespaces. One with a single path to the servers, and one with multiple paths where segment routing was used to determine which path is used for which server.

Tests were realised over the two topologies to identify strengths and weaknesses of the developed load balancing program. The test results show that the program meets the target of balancing the traffic across the different paths and servers in all test cases. They highlight how the need for collecting information about the different client connections influences the accuracy and reactivity of the load balancing. IOAM does not provide information on flow loads on the network. The information was computed by counting the packets each client send to the servers. This is the main weak point of load balancing with IOAM since it introduces a delay and reduces the overall performances. Finally, tests were realised to determine how much IOAM can increase packet loss. The results indicate that including an IOAM header in all packets can generate significant packet losses. The more packets with IOAM header, the higher the packet loss, but including IOAM header in only a quarter of the packets does not create a significant increase in packet loss.

In conclusion, the results in this work indicate that IOAM can be used together with segment routing to implement a dynamic load balancing over a virtual network. It is a promising approach, and future works could explore how to improve the load balancing algorithm, and how to collect information about the client connection size in a more efficient and accurate way. This work provides tests of the load balancing program on a small scale, with only a small number of clients and servers, and low bandwidth. It is a proof of concept and a larger work could be initiated with more tests on a larger scale to determine how far such load balancing program could be used in an actual case. Then it will be important to test load balancing with IOAM and segment routing in real world, on a non-virtual topology.