High-Concurrency Stress Testing for AI Infrastructure: Workloads, Bottlenecks, and Failure Semantics

Abstract

Ai Server Stress Testing is advancing through efforts to align performance with evidence quality, resource limits, and transfer across settings. This methodological synthesis evaluates designing stress tests that expose queueing, memory, accelerator, and network failures without confusing load with realism. The source set brings together 1 focal paper with 12 independently retrieved publications whose DOI or publisher records were checked before inclusion. The analysis is organized around workload models, tail latency, resource contention, failure injection, and capacity planning. A central precaution is not to treat headline results as if they shared one denominator, the review compares task scope, modeling premises, and evaluation limits. Across the literature, the recurring conclusion is that advances in AI server stress testing become credible when workload, dependency, and recovery policy are evaluated together and when uncertainty about observability is reported explicitly. The proposed reading joins method selection to use-case risk and reveals common limits on generalization, and proposes a research agenda centered on well-specified controls, robustness tests, and reproducible workflows.