Using machine learning to predict optimal erasure coding policies for object storage system in OpenStack Swift

Abstract

Erasure coding helps to reduce storage overhead and improve fault tolerance. But the procedure to select an appropriate erasure coding policy is complex, which often involves tradeoffs among various metrics such as access latency, recovery behavior, storage efficiency, etc. Generally, these are handled using static or heuristic-based configurations, which can not account for variations in workload. In the industry, service providers like Ceph do benchmarking based on throughput/latency without considering workload diversity. OpenStack Swift, one of the most widely used open-source object storage systems, supports erasure coding, but it allocates policy selection in a manual, static way - without it being workload-aware. To address this problem, this thesis showcases a data-driven performance modeling framework for erasure coding in object storage systems using machine learning. A structured dataset- ABDS-30k was constructed in a controlled execution of varying workload conditions. It was created on a Swift All-In-One (SAIO) testbed under different erasure coding policies with the addition of failure injections. The dataset collects several empirically observed performance metrics such as read and write latency, tail latency, success rate, and reconstruction time in case of disk failures. For the job of selecting the optimal erasure coding policy, we adopt two distinct Machine Learning paradigms - a regression-based approach of performance modeling and a classification-based approach for direct data-driven policy recommendation. For regression, CatBoost, XGBoost, and Random Forest Regression are used to predict target metrics in a given workload context and failure scenario, and the optimal policy is chosen based on a weighted score. In the classification-based approach, CatBoostClassifier, XGBoostClassifier, and Logistic Regression are used to label workload–policy pairs using an oracle cost function, and the most optimal erasure coding policy is directly recommended. Experimental results show that the regression models have moderate prediction errors for latency and recovery metrics because of the inherently noisy and heavy-tailed nature of the system, but they remain effective in optimal policy recommendation by exhibiting top-1 accuracy with 48.16% in XGBoost Regression and a top-3 accuracy 100% in Random Forest Regression model. Also, the regret mean (0.010 - 1.81) and regret median (0.00216 - 0.089) values showed a very low margin of error. Classification-based approach shows comparatively weaker metrics, indicating that it is not optimal for data-driven policy recommendation. Overall, this shows the feasibility of using ML-based performance modeling as opposed to static and heuristic-based policy selection, which can be later used as a foundation for future control-plane automations.