CRITICALITY-AWARE RESOURCE DISTRIBUTION METHOD FOR CONTAINER ORCHESTRATION SYSTEMS

Abstract

This paper proposes an advanced resource distribution method for container orchestration systems that incorporates service criticality as a key parameter for placement decisions. Unlike traditional scheduling strategies such as “binpack” or “spread”, the proposed approach integrates the business importance of services to mitigate security risks associated with the unwanted colocation of containers with different criticality levels.

The core of the methodology relies on Kohonen self-organizing maps, which is a type of unsupervised artificial neural network. By including a service criticality level into the input vector alongside hardware requirements, the system achieves intelligent clustering of containers on cluster nodes. To evaluate the effectiveness of this placement, the study introduces a “criticality colocation coefficient”. This metric quantifies security risks based on the number of distinct criticality levels present on a single node and the variance between their minimum and maximum values.

The experimental framework included a comparative analysis of the strategy based on self-organizing map both with and without the inclusion of the criticality parameter. The results reveal that while a resource-only self-organizing map configuration offers small improvements over traditional methods, the integration of criticality-awareness leads to a substantial decrease in the colocation coefficient, effectively isolating sensitive services. Furthermore, the study identifies that the topology of the self-organizing map plays a crucial role in the outcome: while a linear topology shows limited effectiveness as the cluster scales, a square matrix topology provides superior results in maintaining clear segregation of critical services.

The findings indicate that the proposed method is highly promising for environments with strict security and stability requirements. Importantly, the addition of the criticality parameter does not lead to significant degradation of other metrics, such as resource defragmentation levels or the success rate of containers distribution. Future research will explore the integration of network interaction patterns and the performance of the model under dynamic workloads, container migrations, and node failures.