POSTER: RDMA Network Performance Anomalies Diagnosis with Hawkeye

RDMA (Remote Direct Memory Access), recognized for its high throughput and low latency, is increasingly deployed across data centers. In RDMA network, the hop-by-hop flow control, i.e., PFC (Priority-based Flow Control), is usually employed to eliminate packet loss and guarantee high performance. However, even with congestion control [9], PFC still frequently occurs, causing head-of-line blocking and congestion spreading [5], which poses new challenges for effective and efficient diagnosis of network performance anomalies (NPAs). For example, in Figure 1, short bursts from A1-An congest SW4.P1 and cause PFC to spread back along the path of flow F2. Finally, SW1.P1 gets PFC and forms into queue congestion. And F1 gets performance anomaly as the victim flow, although it has no queue contention with other flows.