What One Block Reclamation Actually Costs
The cost of garbage collection follows a very simple model. Let u be the fraction of valid pages in the block being reclaimed. Before that block can be erased, those valid pages must be relocated elsewhere, so the net space recovered is 1-u. Each user page written therefore carries roughly u over 1-u in relocation writes, and steady-state write amplification approximates 1 over 1-u. Move u from 0.5 to 0.8 and amplification rises from 2 to 5 — the curve is very steep at high occupancy. That explains two design choices at once: why victim selection favours the block with the fewest valid pages, and why hot/cold separation is worth the effort. Data with similar lifetimes written into the same block tends to invalidate together, driving u naturally toward zero.
Foreground Reads Queue Behind Background Relocation
Write amplification is only the accounting cost; what users feel is resource contention. NAND program and erase operations occupy a die for a long time, and a foreground read that lands on that die simply waits, pushing tail latency up. Controller-side responses include scheduling reclamation at die and plane granularity, using program suspend so urgent reads can cut in, applying a dynamically adjustable bandwidth quota to the reclamation path, and reclaiming ahead of demand during idle windows so a burst of foreground writes does not collide with emergency collection. The low-latency QoS baseline, in implementation terms, is exactly this combination.
Over-provisioning Is the Main Dial
Reserved space sets the average valid page ratio of blocks at steady state. More over-provisioning spreads the same user write volume across more physical blocks, lowering u and improving write amplification along the non-linear curve above. The CORE 7 capacity ladder shows this plainly: the J750 is planned at 1.92, 3.84, 7.68 and 15.36TB, the J770 at 1.6, 3.2, 6.4 and 12.8TB. On the same generation platform the latter trades lower usable capacity for a larger reserve, and its endurance target rises from 1 DWPD to 3 DWPD while the 4KB random write target rises from 300 KIOPS to 600 KIOPS. Endurance and steady-state random write improve together because one variable is driving both.
Why Steady-state Numbers Must Follow Preconditioning
On a fresh drive every block is empty, writes land straight into free blocks and collection barely triggers, so the random write figure measured there approaches the theoretical ceiling of media and channels and says nothing about what the device sustains in long-term operation. Only once every physical block has been written at least once and reclamation is running continuously does write amplification converge to a stable value matched to the reserve and the workload shape. Random write specifications must therefore state that they are given on a steady-state basis, the test flow must include both purge and preconditioning, and the time taken to reach steady state must be recorded. The figures above are planning targets.