Come Hell or Still Water: Alleviating Tail Latency in Cloud Block Store

Maintaining low tail latency is crucial for cloud storage services. In ALIBABA CLOUD, our Elastic Block Storage (EBS), like many others, adopts layers of load balancing to avoid hot-spot I/Os, a dominant contributor to tail latency.

However, in the field, EBS has still been suffering from tail latency spikes. Through extensive analysis of production workloads, we have identified the root cause: the workload bursts caused by a small group of Virtual Disks (VDs), which fundamentally influence the tail latency of the entire cluster. We hence propose a lightweight dual-bucket throttling mechanism to effectively mitigate the issue while maintaining fairness. In addition, we discover that, even under underloaded scenarios, the tail latency remains suboptimal due to the event-loop thread model. We propose a prioritybased scheduling mechanism to separate I/O-related tasks from I/O-unrelated ones. Our evaluation shows that the proposed mechanisms can reduce the tail latency by up to 97% in burst and 43% in underloaded scenarios. Our mechanisms have been deployed across dozens of clusters for more than three months, and have served hundreds of trillions of I/O requests. They reduce the P99999 tail latency of steady segments by 59.7% under burst scenarios and of all I/Os by 22% in underloaded scenarios.

• To avoid ambiguity, throughout this paper, we use the term node (e.g., compute and storage node) to refer to the physical machine, and server to the software process (e.g., proxy server) running on the node.