Improving Paravirtualization using eBPF

Virtual machines (VMs) enabled cloud computing but their isolation prevents software performance hints from reaching hardware, leading to unmet guarantees (e.g., CPU isolation). This creates an inconsistency between software expectations and reality. This thesis addresses this via paravirtualization, allowing host-guest communication. Instead of risky kernel modifications, it introduces a framework injecting eBPF (extended Berkeley Packet Filter) programs into guest VMs. These secure, verified programs provide the host with operational hints without altering the guest kernel. Communication leverages an improved virtio-vsock. An innovative application is also demonstrated: using eBPF to move traffic classification to the VM, significantly increasing packet scheduling throughput in data centers. The research ultimately showcases how guest cooperation, facilitated by eBPF, can enhance cloud computing performance.