Borg’s Resource Management, Isolation, and Lessons for Cloud‑Native Systems

The article provides an in‑depth analysis of Borg, Google's large‑scale cluster manager, focusing on how it achieves high machine utilization through techniques such as cell compaction, workload segregation, and fine‑grained resource requests.

It presents experimental evaluations that compare shared and dedicated Borg cells, measure CPU cycle interference, and assess the impact of resource bucketing on overall machine requirements.

Resource reclamation mechanisms, including dynamic reservation adjustments and aggressive reclamation policies, are discussed, showing how they reduce wasted capacity while maintaining acceptable OOM rates.

The paper also explores isolation strategies—both security (chroot, cgroups, SSH isolation) and performance (CPU pinning, CFS bandwidth control)—and how they mitigate interference between latency‑sensitive and batch workloads.

Finally, the authors summarize lessons learned from a decade of production experience, highlighting design choices that proved robust (e.g., alloc abstraction, introspection tooling) and those that required redesign in newer systems like Kubernetes.