Soft vs Hard Rate Limits

Overview

Soft vs hard rate limits are the two rate-limiting modes in SSD QoS, exposed as a toggle in NVMe TP4176. Hard rate limits strictly cap a tenant at its configured bandwidth/IOPS ceiling regardless of whether other tenants are idle — ensuring strict isolation but potentially leaving overall SSD capacity unused. Soft rate limits let a tenant burst above its allocation to absorb idle capacity left by other tenants, maximizing utilization while still favoring each tenant's baseline allocation.

Concrete examples on a 100 MB/s SSD with three tenants each set to 50 MB/s illustrate the difference. Under hard limits, tenant 3 is capped at 50 MB/s even when tenants 1 and 2 use only 20 MB/s combined. Under soft limits, tenant 3 flows up to consume 80 MB/s by absorbing the idle capacity. Soft limits also power a live-migration use case: a parent process set to a 1 MB/s soft limit can consume the full 100 MB/s SSD when VM tenants go idle, accelerating migration convergence. When an SSD is over-prescribed — for example, three tenants each set to 50 MB/s on a 100 MB/s drive, all active at high queue depth — the tenants converge to roughly equal sharing at about 33 MB/s each; over-prescription is generally not recommended.