Performance Analysis of Switch Workload Management Protocols for Burst Tolerant Networks

One of the main causes of packet loss and performance degradation in today's datacenter operations, is represented by microsecond-scaled congestion events, known as microbursts. Microbursts are characterized by sudden spikes in network traffic, and are likely to lead to congestion, impacting the overall efficiency of the datacenter. Existing solutions, such as packet deflection techniques, have shown promise in mitigating microburst effects. However, further research is needed to explore alternative approaches that can enhance network performance and stability. In this study, we propose Robinhood, a solution aimed at mitigating microbursts in datacenter networks as well as improving the overall performance. We exploit the insights gained from existing solutions, taking inspiration from the job-scheduling domain, and implement novel algorithms based on work-stealing scheduling techniques to address the challenge of microburst mitigation. We implemented a simulation framework from scratch, featuring BRITE for network topology generation, to test different networks under various degrees of load. Through extensive simulations and performance evaluations, we demonstrate the effectiveness of our proposed work-stealing scheduling algorithms in mitigating microbursts: for example, in a leaf-spine architecture network under 80\% of load, the flow completion time improves by 22%, 6% and 7% when applying work-stealing techniques to switch buffers, compared to ECMP, DIBS and Vertigo respectively. Additionally, the flexibility of these algorithms, makes them suitable for applicability to other network environments beyond datacenters, such as 5G and other burst tolerant networks. For future research purposes, we released the source code of the simulation framework, in order to be tested with different network topologies and configurations.