Distributed Cloud Performance Benchmarking on Google Cloud Platform Using MobileNetV2 Inference Workloads

Abstract

This study benchmarked MobileNetV2 inference on Google Kubernetes Engine to examine how cluster configuration and concurrent-user load affected service performance. A 3×3 factorial design compared static two-node, static four-node, and autoscaling two-to-four-node deployments at 10, 50, and 100 concurrent users. Five 15-minute runs per treatment yielded 45 experimental units. Client-side throughput, average latency, and error rate were interpreted alongside CPU and pod telemetry. At 100 users, the two-node configuration recorded 14.40 requests/s, 7,121.81 ms latency, and 70.67% errors. Four static nodes recorded 19.62 requests/s, 4,636.13 ms latency, and 50.53% errors. The 36.25% throughput increase was less than proportional to the doubled node count at that load. Autoscaling recorded 19.82 requests/s and 69.79% errors at 100 users, while its 50-user treatment coincided with pod thrashing. The reported two-way analysis of variance indicated significant configuration, load, and interaction effects at the 0.05 level; exact assumption-test outputs and sums of squares were unavailable for independent verification. The results showed that the tested configurations did not ensure reliable service under high concurrency. Resource limits, timeouts, and autoscaling settings remained plausible contributors requiring controlled follow-up experiments.