Simulation Modeling of a Fault-Tolerant Computing Cluster with Two-Level Load Balancing and Container Virtualization

Abstract

This paper investigates the fault tolerance of a computing cluster with container virtualization and a dedicated load balancer. A discrete-event simulation model is proposed and implemented in Python/SimPy, incorporating the load balancer as a vulnerable single point of failure (SPOF, Level 0) and a pool of worker nodes with containers (Level 1). Unlike classical queueing models, each server undergoes a realistic cascade of hardware/software aging UP → DEGRADE → DOWN, with a host-node failure physically blocking its containers and their migration. Container service rate is dynamically derived from an experimentally measured performance matrix μ(n, m). To capture the true delay profile, the architecture employs infinite persistent queues combined with strict rejection of new requests during hardware downtimes. The software implementation is verified against the analytical M/M/c model in a no-failure regime, yielding a relative error of mean waiting time below 6 %. Tail characteristics are computed over the pooled sample of all runs (pooled metric), eliminating the systematic bias introduced by averaging per-run percentiles. Four series of experiments are conducted (10 Monte Carlo runs each): influence of MTTFfail with three cascade-aging profiles p ∈ {0.7, 0.8, 0.9}, MTTR, balancer routing policies, and arrival intensity λ ∈ [10, 30] tasks/s. Key result: accounting for hidden SPOF-node degradation critically shifts the cluster stability boundary; a saturation threshold λ* ≈ 22–26 tasks/s is observed, beyond which the pooled 95-th percentile of waiting time W₀,₉₅ grows from 0.11 s at λ = 10 tasks/s to 109 s at λ = 30 tasks/s. The model source code and experiments are published in an open repository. The proposed model and the obtained data form a foundation for the synthesis of predictive routing policies and proactive cluster resource management.