A one-year measurement of quota-induced waste in private clouds
This paper analyzes a year of private cloud operational data to reveal that quota-based resource management, combined with user behavior of provisioning for peak demand and administrative overselling, leads to significant inefficiencies characterized by persistent quota exhaustion despite available capacity and prolonged VM idleness.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the world of computing, there is a constant tension between having enough power to do a job and wasting that power when it sits idle. Public cloud services, like those used by major corporations, often solve this by charging users for every second of computer time they consume. This price tag acts as a natural brake; if a server is not doing useful work, the user stops paying for it and shuts it down. Private clouds, however, operate differently. These are dedicated computing environments owned by universities, research labs, or large companies, where users do not pay a bill for the resources they take. Instead, administrators set a fixed limit, known as a quota, on how much computing power each team can reserve. The idea is simple: give every team a fair share of the total capacity so that one group does not swallow everything. But without a price to discourage waste, a strange dynamic can emerge. Teams might grab as much computing power as they are allowed, just in case they need it later, leaving that power sitting empty while other teams struggle to find any space at all.
Researchers at the Federal University of Campina Grande in Brazil and the University of Piemonte Orientale in Italy decided to investigate this invisible waste. They spent a full year, from May 2024 to May 2025, watching the digital heartbeat of a private cloud used by their own university. This system, built on open-source software called OpenStack, serves hundreds of researchers and students. It is a massive machine, capable of running over a thousand virtual computers simultaneously, with a total memory capacity of 1,024 gigabytes. The team did not just look at how much power was being used; they looked at how much power was reserved versus how much was actually doing work. They collected over 64 million data points, tracking every virtual machine, every project, and every five-minute interval of activity. Their goal was to see if the rules meant to keep things fair were actually causing the system to become inefficient.
What they found was a system caught in a paradox of its own making. The cloud was designed with a safety margin, meaning the total amount of computing power promised to all the teams combined was actually larger than the physical machine could hold at once. This is a common practice called overselling, based on the hope that not everyone will need their full share at the exact same moment. In reality, the physical machine rarely reached its limit; it was completely full on only about 3 percent of the days. Yet, despite this abundance of empty space, many teams were constantly blocked from creating new virtual machines. The researchers discovered that at least half of the projects hit a wall where they could not provision even the smallest possible computer because they had run out of their personal quota allowance. About 14 percent of these projects were stuck in this state for their entire time in the study. Even more telling, roughly 27 percent of the projects were blocked from adding new machines even when the entire cluster was using only 70 to 80 percent of its total physical memory. The bottleneck was not a lack of hardware; it was the rigid rules of the quota system itself.
The reason for this blockage lies in how the users behave when there is no cost to holding onto resources. The study revealed a pattern of resource hoarding. When a team receives a quota, they tend to claim their full share immediately, often provisioning virtual machines long before they actually need them. These machines then sit idle, consuming their share of the quota, while the team waits for a future project to begin. The researchers measured this by looking at how much time a virtual machine spent actually working versus how much time it sat doing nothing. They defined "meaningful use" as times when the processor was actively calculating, ignoring times when the machine was just holding data in memory. The results were stark: across the entire system, virtual machines spent more than twice as much time idle as they did in meaningful use. While about half of the machines started working on the very day they were created, a significant number sat completely unused for weeks or even months.
This behavior creates a vicious cycle. Because users cannot be sure they will get resources back later if they give them up, they hold onto them tightly. They create a virtual machine today, leave it running with nothing on it, and keep it there for weeks, just to ensure it is available when a deadline arrives. This is a rational choice for the individual team, but it is disastrous for the system as a whole. It means that while one team is sitting on a pile of unused computing power, another team is unable to start a new project because their quota is full. The researchers noted that this is not a failure of the hardware or a lack of demand; it is a failure of the incentive structure. Without a price to pay for idle time, and without a way to guarantee future access through a reservation system, the only safe strategy for a user is to hoard.
The study suggests that the solution is not simply to build bigger machines or give out more quotas, as that would likely just lead to more hoarding. Instead, the administrators need to change the rules of the game. They could introduce mechanisms that allow users to reserve resources for a specific future time, giving them the security they currently seek without forcing them to hold onto idle machines. They could also implement systems that detect when a machine has been sitting unused for too long and automatically reclaim that space for others. The data shows that a huge amount of potential computing power is being wasted, not because it is broken, but because the people using it are playing a game where the rules encourage them to grab and keep. By understanding this behavior, the researchers hope to design private clouds that are not just fair in theory, but efficient in practice, ensuring that the powerful machines built for research are actually used for research.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.