Two Kinds of Shingled Drive

A conventional (CMR) hard drive writes each track separately, with a gap between neighbours. An SMR drive overlaps tracks like roof shingles. The write head is wider than the read head, so it can pack tracks closer together and gain capacity on the same platters. The price is that you can’t overwrite one track in place without clobbering the ones next to it. Data has to be written in order, in groups of tracks called zones.

What differs between the two kinds of SMR is who deals with that restriction.

Drive-managed (DM-SMR)

The drive hides it from you

Firmware presents a normal random-write disk. Behind the scenes it stages writes in a CMR cache area and rewrites shingled zones later. It works with any OS, but you have no control over when the drive gets busy, and capacities top out around 8 TB.

Host-managed (HM-SMR)

The host follows the rules

The drive exposes its zones and refuses out-of-order writes to sequential zones. Software must write at each zone’s write pointer. There is no hidden translation layer, and capacities run to 24 TB and beyond.

That capacity gap is the main reason host-managed drives matter. Drive-managed models never made it past the single-digit terabytes, while the high-capacity drives built for datacentres and archives are host-managed.

Writing From LBA 0, Forward Only

Most storage software sits on top of a filesystem, and filesystems are full of small in-place updates: journals, allocation bitmaps, inode tables, directory blocks, atime updates. Each is a tiny random write. On a CMR disk that costs a seek. On an SMR disk it can mean read-modify-write across an entire shingled band, or, on a host-managed drive, an outright error.

HuskHoard sidesteps the problem by writing directly to the disk with no filesystem in the way. A volume starts at LBA 0, and every file is written sequentially after the one before it. There is no journal to replay, no cache to flush, and no metadata being rewritten near the data. The state of the volume lives in the Husk Catalog on the host, not scattered across the disk.

Why this matters

Sequential, append-only writing is the one workload SMR was designed for. We didn’t bolt SMR support onto HuskHoard. The write model we chose for archives happens to be what shingled drives want.

Drive-Managed: Staying Out of the Drive’s Way

The notorious DM-SMR failure mode goes like this. A filesystem sprays random writes at the drive. The drive absorbs them into its CMR cache until the cache fills. Then it has to clean up, rewriting shingled zones in the background, and write speeds can collapse from over 100 MB/s to a trickle. RAID rebuilds and ZFS resilvers on these drives are the classic horror stories.

With HuskHoard that situation doesn’t arise. Because writes are strictly sequential from LBA 0, the drive’s firmware sees a clean stream of forward writes. There is no random traffic to stage in the cache and no backlog of zones waiting to be cleaned. The drive behaves like what it physically is: a large, sequential-write device.

Host-Managed: Where the Capacity Is

Host-managed drives drop the pretence. They divide the disk into zones, typically 256 MB each, and most of them are sequential-write-required. Each has a write pointer, and a write that doesn’t land exactly on the pointer is rejected. To reuse a zone, the host explicitly resets its pointer. A small number of conventional zones may exist for metadata, but an append-only design doesn’t need them.

Since we never go backwards, that contract costs us nothing. Linux sees these drives as zoned block devices, and you can inspect one directly:

# Is this drive host-managed? $ cat /sys/block/sdb/queue/zoned host-managed # Look at the first few zones $ blkzone report /dev/sdb | head -3 start: 0x000000000, len 0x080000, cap 0x080000, wptr 0x080000 reset:0 non-seq:0, zcond:14(f) [type: 2(SEQ_WRITE_REQUIRED)] start: 0x000080000, len 0x080000, cap 0x080000, wptr 0x080000 reset:0 non-seq:0, zcond:14(f) [type: 2(SEQ_WRITE_REQUIRED)] start: 0x000100000, len 0x080000, cap 0x080000, wptr 0x012c00 reset:0 non-seq:0, zcond: 2(oi) [type: 2(SEQ_WRITE_REQUIRED)]

Two zones are full, the third is partly written, and everything after it is empty. That is exactly the picture an append-only archive should produce. The advantage over drive-managed is that nothing is hidden. There’s no surprise garbage collection, because the drive never does any.

Connecting Host-Managed Drives

This is the part that catches people out. Host-managed drives are controlled with zone commands (REPORT ZONES, RESET WRITE POINTER and friends) sent over the drive’s native SATA command channel. Most USB bridges don’t pass that channel through. Plug a host-managed drive into a typical USB enclosure and the bridge chip hides the zone commands, so the drive either shows up as an unusable disk or doesn’t work at all.

✓
SATA on an HBA
The simplest and most reliable option. Connect the drive to an HBA or motherboard SATA port, and the full command set reaches the drive.
✓
USB4 / Thunderbolt only
If you need an external connection, use USB4 (Thunderbolt) enclosures that expose a real SATA or SAS controller to the host. Don’t use standard USB 3.x bridge enclosures.
✗
Standard USB 3.x enclosures
Fine for drive-managed disks. Not for host-managed: the zone commands get dropped.

When in doubt, run the blkzone check above. If it can’t report zones, the connection is the problem, not the drive.

Cost and Power Efficiency

Shingling squeezes more bits onto the same platters, and that carries through to both the price and the electricity bill.

01 — Cost per TB
More Capacity per Platter
The same heads and platters hold more data, so the cost per terabyte falls. Savings against equivalent CMR drives are typically in the range of 10 to 20%, and the gap tends to be widest at the highest capacities.
02 — Power per TB
Fewer Spindles
Drive power draw is dominated by spinning platters and moving heads, and it doesn’t scale with capacity. A 24 TB drive uses far less power per terabyte than three 8 TB drives. Fewer drives also means less cooling and fewer failure points.
03 — Idle Cost
Spin Down Between Jobs
An archive is mostly idle. Tiered, write-once data can sit on drives that are spun down or powered off, and a drive on a shelf draws nothing. The catalog keeps every file searchable meanwhile.
04 — Fit
No Penalty for Our Workload
SMR’s weakness is random rewrites. An append-only archive never does them, so you get the density and price advantage without paying the performance tax most people associate with SMR.

Exact prices and wattages vary by model and market, so check the datasheets for the drives you’re considering. The direction is consistent, though: for cold, write-once data, SMR delivers the lowest cost and power per terabyte of any spinning disk.

Which One Should You Buy?

SMR isn’t the enemy of your data. A random-write workload on a drive that can’t take one is. Give shingled drives the sequential stream they were built for, and they’re some of the best-value archive media you can buy.