Classification of load balancing in the internet

• Autor(a) principal: Rafael Luis Caldas Almeida
• Data de Publicação: 2019
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Repositório Institucional da UFMG
• Texto Completo: http://hdl.handle.net/1843/34630
• topic: Internet; Routing; Protocols; Computer networking; Computação – Teses.; Redes de computadores - Teses.; Internet – Protocolos – Teses.; IPv6 – Teses

A router may perform load balancing and distribute traffic across multiple routes that have the same cost. Load balancing improves available bandwidth, robustness to failures, and performance. Routers that perform load balancing (referred to as load balancers) compute the link a packet should be forwarded to as a function of the packet’s flow identifier, a subset of fields in the packet’s headers (e.g., IP addresses and port numbers). Network operators and researchers rely on measurement tools to identify and characterize load balancing. However, recent advances in programmable data planes, software defined networks, and even the adoption of IPv6, support novel, more complex load balancing strategies. These strategies allow the definition of flow identifiers that existing measurement tools are incompatible with. In this work, we introduce the Multipath Classification Algorithm (MCA). We generalize the network formalism used to describe load balancing and extend existing techniques to consider that load balancers may use arbitrary combinations of packet header fields for load balancing. MCA detects load balancers that existing tools cannot, regardless of the bits load balancers consider in flow identifiers. Furthermore, MCA classifies the behavior of load balancers and their impact on application traffic. We propose optimizations that reduce the classification cost by 11% and the overall cost by 6%, without loss of accuracy. Our evaluation shows that the process of classifying load balancers entails a cost similar to the cost of the detection process, demonstrating MCA is a practical tool. Finally, we use MCA to collect a representative dataset of route measurements to characterize load balancing in the Internet. Our results show that load balancing is more prevalent and load balancing strategies are more mature than previous characterizations have found.