In short
This study proposes a framework based on conformal prediction for selecting the optimal size of virtual machines in the cloud. The approach accounts for workload unpredictability and reduces the risks of over- and under-provisioning.
Traditional algorithms for allocating virtual machines in the cloud do not handle load fluctuations well. The result is chronic over-provisioning, where you pay for unused resources, or under-provisioning, which leads to degraded performance. A paper on arXiv proposes solving the problem of VM sizing (right-sizing recommendations, RSR) using interval forecasting based on conformal prediction.
The key idea is not to predict a specific utilization value, but to construct confidence intervals that accurately reflect the uncertainty in resource demand. To achieve this, a bootstrap version of conformal prediction is used, applied to load patterns and correlations between time series.
The authors state that ML regression models evaluated via backtesting yield “promising” results for utilization forecasting. They also propose a model ranking for long-lived VMs. The abstract does not include specific benchmarks or savings figures, so the scale of the gains remains unclear.
Practical takeaway: Right-sizing via conformal prediction is a viable approach for cloud operators who need to automate capacity planning. However, there is still a significant gap between the research framework and the production pipeline, and without publicly available metrics based on real-world data from hyperscalers, it is premature to judge the actual cost savings.
Source: cs.AI updates on arXiv.org