GraphFlow: A Container-Based Partitioned Global Address Space Framework for Distributed Graph Analysis on Kubernetes

Distributed graph analytics increasingly rely on scalable communication models and dynamic resource management to meet the demands of modern data-intensive applications. While partitioned global address space programming models, such as UPC++, provide an efficient and expressive substrate for irregular, communication-heavy workloads, their integration into cloud-native environments has received limited attention so far. This work presents GraphFlow, a container-based PGAS execution framework that empowers high-performance graph processing on Kubernetes clusters. GraphFlow integrates a custom UPC++ Operator, a deterministic master–worker pod topology manager, and a REST-driven job orchestration layer, which collectively enable portable, elastic, and programmable execution of distributed partitioned global address space applications. The framework decouples resource allocation from job execution, simplifies operator-level control, and exposes a uniform interface for launching, monitoring, and terminating partitioned global address space workloads. An experimental evaluation using representative graph kernels shows that containerization introduces only marginal overhead compared to native execution while offering significant gains in automation, reproducibility, and scalability management. GraphFlow illustrates the feasibility of extending cloud-native orchestration tools to support high-performance partitioned global address space workloads and lays the foundation for future research on elastic, fault-tolerant, and serverless graph analytics.