Daredevil: Rescue Your Flash Storage from Inflexible Kernel Storage Stack
Published in ACM European Conference on Computer Systems (Eurosys), 2025
Existing kernel storage stacks struggle to mitigate the multi-tenancy issue for NVMe SSDs, due to performance interference between I/O requests from tenants with different SLAs. Addressing this requires separating their I/O requests within the NVMe I/O queues (NQs). However, our analysis reveals that the static CPU core-NQ bindings of current storage stacks restrict their flexibility to achieve this goal.
We propose Daredevil, a novel kernel storage stack, which addresses this issue by decoupling the static bindings and allowing full connectivity between cores and NQs. Therefore, it grants multi-tenancy control the flexibility to freely route requests among NQs according to their SLAs. Moreover, it incorporates multi-tenancy-aware scheduling on NQs to facilitate efficient request routing. Our evaluation shows that Daredevil can reduce I/O request latency by up to 3-170× compared to current kernel storage stacks, while maintaining comparable throughput.
Recommended citation: N/A.