Experimental evaluation of the accuracy of empirical metrics for identifying the leading load scheduling algorithm when accessing external web services
Semen Antonovich Esaev, Aleksei Nikolaevich Alpatov · Программные системы и вычислительные методы · 2026
The subject of the study is the process of adaptive load scheduling when accessing external web services with fixed request limits. In large microservice architectures, the role of data aggregation is usually assigned to a separate service. Such a service is required to comply with the limitations of external integrations and process incoming requests with minimal waiting time. Scheduling request execution using a single algorithm turns out to be impractical if the characteristics of the input flow change during operation, so the aggregation service needs to be able to assess the state of the input flow in order to switch to the algorithm most suitable for these conditions. The purpose of the work is to assess the sufficiency of empirical metrics of the input flow state to determine the required scheduling algorithm in conditions where there is no accurate analytical model of the system, and the algorithm itself affects the observed values of metrics. To achieve this goal, a simulation environment was implemented, synthetic data was prepared, the performance boundaries of the considered algorithms under stationary conditions were investigated, and empirical metrics were proposed and investigated: the load factor and the coefficient of variation in the number of requests per task. Based on the research results, a map of the dominance of algorithms and a map of the observed metrics have been compiled. The scientific novelty lies in the experimental assessment of the suitability of a pair of empirical metrics for determining the leading scheduling algorithm based on the dominance map in a system with a limited request frequency. The study showed that the implementation of metrics independent of algorithms, even with a weak correlation with the theoretical value, makes it possible to determine the leading algorithm with an accuracy of 92.8%, while errors are localized in the transition zone between the areas of dominance. The accuracy of the determination in the zones with a statistically significant leader was 96.6%, in the transition zones – 86.1%. Analysis using the Student's paired t-test showed that in 63.8% of the map cells, the dominance of the leader is statistically significant (p " 0.05), and the remaining cells correspond to transition zones in which the algorithms show comparable results. The considered technique can be used in an industrial environment to adjust adaptive switching thresholds. A further direction of research is the study of the behavior of metrics in dynamic conditions when changing the characteristics of the input flow.