Detectando Heavy Hitters globalmente em dispositivos programáveis multi-pipes

Abstract

High-precision monitoring mechanisms play a fundamental role in managing various network tasks, such as congestion control, anomaly detection, load balancing, among others. One way to contribute to network management is by applying a detection of flows that have a greater influence on traffic, characterized as HH. HH flows correspond to flows that account for the largest share of transmitted bytes across the network, consequently consuming more resources. By utilizing programmable hardware (Switches, SmartNICs, and DPUs) in conjunction with programming languages like P4, it is possible to detect these flows at in-line rate, meaning detection directly in the network’s data plane. The literature reveals that HH flow detection is a widely explored topic. However, studies proposing solutions for HH detection erroneously assume that the switch has only a single pipe. In reality, programmable switches have multiple pipes, ranging from 2 to 16. In light of this, this study presents an approach to identifying HH flows in programmable switches with multiple pipes. The development encompasses two detection approaches. In the first approach, an accumulator located in the switch centralizes data from all pipes and communicates with the control plane. In the second approach, communications with the control plane are independent for each pipe. Our approaches were validated through an emulator, which allowed us to obtain preliminary results indicating the effectiveness and improvement in HH detection in multi-pipe switches, compared to detection in switches operating with a single pipe.