Clouded Comparisons - On the Impact of Virtual Machines on TCP-BBR Performance

Recently, TCP performance studies [1, 3, 4, 8] have focused on BBR which estimates the available bandwidth based on packet delay - instead of strictly reacting to packet loss as a sign of congestion. In contrast to packet loss, delay is more susceptible to jitter induced by the various system components in experimental setups and real-world deployments, including virtualization. Considering that BBR deployments in the wild are often running in Cloud-based virtualized environments, we measure how the choice of virtualization affects BBR v1-3 throughput in different BBR-enabled Linux kernels. Our results reveal that when using Google's custom kernels - the only publicly available source of BBRv2/v3 - BBR performance deteriorates under virtualized scheduling conditions, reducing throughput to almost zero in VirtualBox-based VMs. Overall, our work raises questions about the robustness of BBR under certain sub-optimal scheduling conditions.