Flash translation, write amplification, and honest persistence

A flash translation layer maps host logical addresses to physical flash locations. Updating a logical block can write a fresh physical location and invalidate the old one. To reclaim an erase unit containing both live and obsolete data, garbage collection may copy the live portion elsewhere before erasing it. Those copies add physical writes the application never explicitly issued. The original SSD design literature explains these mechanisms, but its historical chip sizes and timings are not specifications for a current device; separate the durable architectural idea from obsolete numerical examples.

Wear leveling tries to avoid exhausting a small subset of erase units while others remain lightly used. Extra movement, available spare capacity, the hot/cold data mix, device firmware, and workload history influence amplification and latency. Discard or deallocation can tell a device that certain logical contents are no longer needed, helping avoid some unnecessary preservation, but timing and exact behavior depend on the interface and implementation. A filesystem delete therefore should not be equated with immediate physical erasure or with a security-grade sanitization operation.

A simple budget divides an assumed remaining physical-write allowance by physical bytes written per day. It can illuminate why reducing redundant rewrites matters, but it is not a predicted failure date. Real media can fail for other reasons, vendor endurance ratings describe stated workloads and conditions, and exposed counters may count host writes rather than actual NAND writes. Name every numerator and denominator. For AI infrastructure, compare full checkpoint rewrites, incremental formats, optimizer-state volume, and temporary files before trying to infer device health from training step count alone.

Benchmark sustained writes after a controlled preparation period as well as short bursts. A device can temporarily use a faster region or available free space, then behave differently while reclaiming blocks. Garbage collection and thermal or shared-resource effects can produce tail-latency changes that an average hides. Keep temperature, capacity utilization, firmware, workload pattern, and synchronization mode with the result. Avoid extrapolating a sequential empty-device throughput number to small durable random writes or to a nearly full drive receiving concurrent checkpoints.

A device with volatile write-back caching may acknowledge an ordinary write before its data is nonvolatile. Linux's block interface distinguishes a preflush, which establishes persistence of previously completed writes before a dependent operation, from force unit access, which requires the flagged write to be nonvolatile at its completion. FUA is not by itself an ordering guarantee for unrelated earlier writes. Filesystems and block layers compose these mechanisms; remapping layers, virtual disks, controllers, and devices must preserve the required semantics. Documented power-loss protection may change which cache is effectively persistent, but the label SSD alone proves none of this.

The queue lab illustrates the demand side only, using equal-sized host-equivalent work units under a fixed conversion. Press Reset, set both arrival to two and effective service to three, predict drainage from four units, then Play or Step. For a separate hypothetical higher-amplification case, Reset, re-enter arrival two and service one, predict growth from four, then Play or Step. Each starts afresh: a rate edit retains elapsed time and recalculates from the initial four, rather than preserving the previous backlog. Lower effective service represents more physical work per host write; the model does not reproduce flash translation, garbage collection, wear, flush, or FUA.