Scheduling Cloud Block Storage Proactively and Reactively with Omar

We explore the performance improvements for a production Elastic Block Storage (EBS) service at Alibaba, which serves millions of users and virtual disks daily. Despite various load balancing measures, Alibaba EBS still faces frequent performance variations or even occasional Service Level Objective (SLO) violations. Our trace analysis reveals that the existing scheduling mechanism in Alibaba EBS, which adopts a common practice of focusing on load balancing alone, is insufficient for mitigating tail latency due to negligence of correlated I/O workloads, reliance on fixed scheduling frequencies, and insufficient input metrics for scheduling decisions. We present Omar, a novel hybrid proactive and reactive scheduling framework that enables dynamic, finegrained scheduling. Evaluation on production-scale testbeds shows that Omar reduces 64 KiB tail latency by up to 52.2% and the number of scheduling operations by 44.6% compared to the baseline scheduler, while incurring limited overhead.