Plan the box

Sizing calculator

We measured each constant below on real IBM i saves to OpenVTL™ appliances; the notes at the bottom say where. The model interpolates the measured retention curves at your dedupe granularity, and where it runs past the data it extends the last measured marginal.

Workload

Full save size (GiB, raw as the host writes it)

Retained full saves (generations kept on the VTL)

Data changed per day (GiB, new or modified between saves)

Save style

Appliance RAM (GiB, any amount, it's a VM)

RAM sets the dedupe granularity at pool creation, and results step at those thresholds: 24+ GiB → 16K · 12–23 → 32K · 6–11 → 64K · under 6 → 128K. Each tier uses its own measured dedupe curve; finer granularity catches more. The appliance reads kernel-visible memory, which runs ~0.5–1 GiB under what you provision, so give a threshold a margin: 25 GiB lands 16K; exactly 24 may not.

Recommendation

dedupe granularity (recordsize)
first full on disk
each additional full ≈
total for retention
effective reduction
data disk (with 30% headroom)
dedupe SSD

Where the numbers come from

Measured through 2026-07 on IBM i POWER over 8 Gb FC, ZFS zstd + Fast Dedup. Compression: 1.45× on raw save streams (host-side DTACPR is irrelevant; the VTL declines drive compression, so streams arrive raw). Retention: 30 identical full-system saves of a ~250 GiB system landed in 577 GiB at 16K granularity,12–13× end-to-end, with the marginal cost converging to the true inter-save churn. Granularity: an offline block-hash sweep of that same dataset, plus 10 identical library saves, measured the complete dedupe-versus-generations curve atevery recordsize tier. Full-system dedupe roughly halves per recordsize doubling (9.1× at 16K down to 1.9× at 128K over 30 generations, because the save engine shifts content in ~16K units); library saves dedupe well at every tier (9.5× to 6.7× over 10 copies). The calculator interpolates those measured curves directly. The dedupe table lives on a dedicated SSD (~27 GiB per 500 GiB of unique data at 16K; coarser records shrink it) and is capacity-bounded; overflow degrades gracefully, never thrashes. Conservative by design: the model predicts 589 GiB for the study that measured 577, and takes no credit for dedupe within a single save.