• Duration: September 2021 to June 2022
• Location: EPFL

Motivation

Linux cgroup (control group), the Linux process and resource management subsystem, is an essential component of container, Kubernetes and Serverless environments. We try to figure out the cgroup control-plane scalability issues.

Overview

Used eBPF to measure the lock contention of cgroup control plane. Modified the kernel to mitigate the lock contention by changing a global mutex as multiple reader-writer locks. The modified cgroup scaled on a 224-core server.

Related Techniques

• Linux Kernel Programming
• eBPF: Monitoring and Benchmarking
• perf
• C
• Python and Bash Script
• QEMU
• gnuplot

Status

Done.