Interference-Aware Workload Placement for Improving Latency Distribution of Converged HPC/Big Data Cloud Infrastructures

Interference-Aware Workload Placement for Improving Latency Distribution of Converged HPC/Big Data Cloud Infrastructures

연구 분야: Software Development

학회: International Conference on Embedded Computer Systems

Recently, High Performance, Big Data, and Cloud Computing worlds tend to converge in terms of workload deployment with containerization technology acting as an enabler towards this direction. In such cases of application diversity and multi-tenancy, a universal scheduler able to satisfy the end-user needs for seamless, yet, efficient application deployment is required. While Kubernetes container orchestrator seems to be the answer that enables application-agnostic deployment, it still depends highly on coarse system metrics for its scheduling policies, thus, neglecting the performance degradation due to resource contention in the underlying system. In this paper, we design and implement an interference-aware modular framework, able to balance incoming workload based on low-level metrics monitoring. We evaluate our proposed solution over different workload mixes and co-location scenarios showing that against the state-of-art, but interference unaware Kubernetes scheduler the proposed framework significantly improves the latency distribution of the converged cloud infrastructure, improving median latency up to 27% and reducing standard deviation up to 25%.