Abstract & Overview
This paper aims to explore the integration of serverless architectures with Kubernetes and OpenStack to optimize AI workflows. The study evaluates the architecture’s scalability, performance, resource utilization, cost efficiency, and energy consumption while addressing challenges such as cold start latency and GPU resource management.The research involves an experimental setup consisting of OpenStack for IaaS, Kubernetes for container orchestration, and serverless frameworks like Knative and OpenFaaS for function execution. AI workflows, including data preprocessing, model training, and real-time inference, were implemented. Performance metrics such as latency, throughput, energy efficiency, and cost were analyzed using tools like Prometheus, Grafana, and TensorFlow Profiler. Comparisons were drawn between serverless and containerized workflows across diverse workload scenarios. The integration demonstrated significant benefits for AI workflows, particularly in real-time inference tasks, with serverless architectures exhibiting better scalability and cost efficiency. Containerized workflows achieved superior GPU utilization and cost performance for batch processing tasks. Serverless workflows reduced energy consumption by up to 25% during idle periods but were impacted by cold start latencies and resource contention during peak workloads. The findings emphasize the transformative potential of serverless-Kubernetes-OpenStack integration, particularly for scalable and energy-efficient AI workflows. However, trade-offs in performance and architectural complexity were noted. The study contributes by optimizing GPU utilization, reducing energy consumption, and supporting hybrid workloads, addressing key gaps in prior research. Recommendations include strategies for latency mitigation, resource orchestration, and workload placement.