A one-year measurement of quota-induced waste in private clouds

Abstract In contrast to public clouds, where pricing regulates resource access, private clouds rely on administrative constraints such as quotas. Typically, users pay nothing for allocated resources; the quota simply caps how much they can provision. Because quotas are oversold in practice, it is unclear whether physical capacity or the quota itself limits provisioning. We analyze one year of operational data from a private cloud, comprising quota assignments, per-project allocation, and per-virtual-machine (VM) utilization sampled every five minutes, totaling over 64 million records. We find this mechanism prone to inefficiencies. Quota exhaustion is a persistent constraint independent of available capacity: at least half of projects cannot provision even the smallest available VM for some period; about 14% remain in this state throughout their entire observed periods, and approximately 27% remain quota-bound even when aggregate memory allocation is between 70% and 80% of physical capacity. Compounding this, overall VM usage is very low: measured by virtual CPU usage, VMs are idle more than twice as long as they are meaningfully used, and although about 50% reach first meaningful use on the day they are created, a long tail remains unused for weeks. We argue this is not coincidental: users request quotas sized for peak demand, while administrators oversell capacity, betting that not everyone peaks simultaneously. With no guarantee of obtainability, users provision in advance and leave resources idle until needed. We conclude by discussing the incentive structure that plausibly sustains this gap and its implications for quota design in private clouds.