Noisy Neighbor: Exploiting RDMA for Resource Exhaustion Attacks in Containerized Clouds

TL;DR

The paper addresses performance isolation in RDMA-enabled container clouds and shows that adversarial RDMA workloads can exhaust RNIC microarchitectural resources, causing severe degradation of co-resident containers. It analyzes two attack families—state saturation and pipeline saturation—through experiments on NVIDIA BlueField-3 and demonstrates substantial bandwidth loss, latency inflation, and link-level congestion with amplification. To restore predictable security guarantees, it proposes HT-Verbs, a threshold-driven framework that uses real-time per-container RDMA verb telemetry to classify RNIC resources into hot, warm, and cold tiers and throttle abusive workloads without hardware changes. The work highlights risks in multi-tenant container environments and offers a practical defense path that blends telemetry, adaptive thresholds, and inline enforcement to preserve performance isolation.

Abstract

In modern containerized cloud environments, the adoption of RDMA (Remote Direct Memory Access) has expanded to reduce CPU overhead and enable high-performance data exchange. Achieving this requires strong performance isolation to ensure that one container's RDMA workload does not degrade the performance of others, thereby maintaining critical security assurances. However, existing isolation techniques are difficult to apply effectively due to the complexity of microarchitectural resource management within RDMA NICs (RNICs). This paper experimentally analyzes two types of resource exhaustion attacks on NVIDIA BlueField-3: (i) state saturation attacks and (ii) pipeline saturation attacks. Our results show that state saturation attacks can cause up to a 93.9% loss in bandwidth, a 1,117x increase in latency, and a 115% rise in cache misses for victim containers, while pipeline saturation attacks lead to severe link-level congestion and significant amplification, where small verb requests result in disproportionately high resource consumption. To mitigate these threats and restore predictable security assurances, we propose HT-Verbs, a threshold-driven framework based on real-time per-container RDMA verb telemetry and adaptive resource classification that partitions RNIC resources into hot, warm, and cold tiers and throttles abusive workloads without requiring hardware modifications.

Figures (14)

• Figure 1: Overview of RDMA workflow.
• Figure 2: Architecture of an RNIC and an example of SR-IOV-based virtualization to serve multiple containers.
• Figure 3: Example of a performance isolation issue where an attacker exhausts the victim's bandwidth by increasing the number of its own QPs.
• Figure 4: Our threat model, where an attacker container compromises the performance isolation of a target host by leveraging a co-resident decoy container to launch RDMA-based resource exhaustion attacks against victim containers.
• Figure 5: Overview of resource exhaustion attacks in an RDMA-enabled container environment.