Astraea: Enforcing DPU Performance Isolation in Public Clouds

This is the authors' abstract. We don't add key points for this paper.

The absence of effective performance isolation mechanisms hinders deploying Data Processing Units (DPUs) in public clouds for multi-application co-location scenarios. To address this challenge, we introduce Astraea, an efficient DPU performance isolation framework targeting DPU-specific computational resources. Astraea overcomes significant obstacles posed by proprietary, high-level SDK interfaces and coarse-grained First-Come-First-Served (FCFS) scheduling via resource occupation profiling, task splitting and workload-guided scheduling. Our open-source prototype on NVIDIA BlueField-3 DPUs reduces SLA violation rates for latency-sensitive applications from 47.66% to just 14.84% versus natives, while introducing less than 4% performance overhead.