Software engineers who work with storage infrastructure regularly encounter SMR drives in vendor discussions and product comparisons, usually presented as "high-capacity drives with some write limitations." That framing undersells both the benefit and the constraint. SMR drives are physically different from conventional magnetic recording drives in a way that has direct consequences for how the filesystem layer must behave. This post explains the geometry from first principles so that the engineering tradeoffs make sense.
How Conventional Magnetic Recording Works
A conventional hard drive platter stores data as magnetic domains on concentric circular tracks. The read/write head is narrow enough to address a single track at a time. Each track has a guard band on either side: a small unwritten region that prevents magnetic interference between adjacent tracks. These guard bands exist because the write head is slightly wider than the read head; writing to one track inevitably magnetizes a small portion of the adjacent track if they are too close together. Guard bands give you separation so that a write to track N does not corrupt track N+1.
Conventional magnetic recording (CMR) drives have been the standard architecture for decades. Their primary engineering constraint is areal density: how much data per square centimeter of platter surface. Guard bands consume area without storing data, so reducing guard band width increases areal density. But reducing guard bands below a certain threshold causes cross-track interference, which produces read errors. CMR drives are near the physical limits of this trade-off.
The SMR Write Geometry
Shingled magnetic recording changes the relationship between tracks to push past the areal density limits of CMR. Instead of writing tracks with guard bands, SMR writes tracks that overlap the previously-written track, similar to how roofing shingles overlap each other. Each new track partially overwrites the edge of the previous track.
The key insight is that the write head is wider than the read head. When you write track N+1, you write over the outer edge of track N with the SMR write head, but the data on track N that remains is still readable by the narrower read head. The overlapping structure eliminates guard bands between tracks, packing tracks closer together and increasing areal density by roughly 20 to 25 percent compared to CMR drives of equivalent physical dimensions.
This is where the constraint comes from. If you need to rewrite data on track N after track N+1 has been written, you cannot simply overwrite track N in place. The SMR write head is wide enough that writing to track N would partially overwrite track N+1, corrupting its data. To update track N, you must first read the current contents of track N+1, rewrite track N with the updated data, then rewrite track N+1 with its original data. If the tracks are grouped in bands of many tracks, this read-modify-write cascade propagates across all tracks above the one you are updating within the band.
Write Bands and the Unit of Sequential Commitment
SMR drives organize their tracks into write bands, typically containing several hundred to several thousand tracks each, depending on drive geometry and firmware design. The fundamental operational rule for SMR drives is: writes must proceed sequentially within a band, from the lowest track to the highest. You cannot write to an arbitrary track within a band without triggering the read-modify-write cascade for all tracks above it in the band.
Each band has a write pointer, analogous to the write pointer in a log-structured storage system. Writes to a band advance the pointer sequentially. When a band is fully written, it is committed and the write pointer resets. To update data in a previously-written band, the entire band must be erased and rewritten from the beginning, which requires reading all still-valid data, writing it to a temporary location, erasing the band, and rewriting the updated content sequentially from the start.
This is why SMR drives are described as "append-only" at the band level. It is not that you cannot update data; it is that updating data within a committed band is expensive because the band erase and rewrite cycle involves reading potentially hundreds of megabytes of data before you can write even a small update.
How Generic Filesystems Break SMR Drives
Generic filesystems such as ext4 and XFS were designed for CMR drives where writes can be directed to any location on the platter without cascading effects. These filesystems issue small random writes constantly: inode updates, directory entry modifications, journal writes, and data block writes all interleave with no awareness of the SMR band structure.
When these random writes hit an SMR drive, the drive's internal firmware has two options: honor the random-write request by triggering the read-modify-write cascade across the affected band, or buffer the write in a small cache region (the persistent media cache, or PMC) and defer the cascade until the cache fills. Most SMR drives use the second approach, maintaining a CMR-like persistent media cache of a few gigabytes that absorbs random writes. The cache eventually fills and must be drained to the SMR bands through read-modify-write, which can stall write throughput dramatically for several minutes at a time. This is the "drive-managed SMR" or DM-SMR model used in consumer drives.
Host-managed SMR (HM-SMR) drives expose the band structure directly to the host and require the host filesystem to respect sequential write ordering. There is no hidden cache to mask the constraint. These drives are the appropriate choice for high-density storage systems where predictable throughput matters more than drop-in compatibility with CMR workloads.
What Proper SMR Management Looks Like
An SMR-aware filesystem or storage system must maintain per-band write pointers and a coalescing write buffer. Writes from the application layer arrive in arbitrary order and size. The storage software accumulates these writes in memory, aligns them to band boundaries, and flushes them to each band as a single large sequential write. The drive sees only sequential, band-aligned writes and operates at its rated sequential write throughput throughout.
In Leil's write scheduler, incoming writes are sorted by logical offset, grouped into band-sized batches, and issued to the drive in sequential band order. Concurrent writes from multiple clients are coalesced before touching the drive. The application layer sees POSIX write semantics: random-order, arbitrary-offset writes are accepted and acknowledged. The drive sees sequential writes only. The coalescing buffer is sized to absorb the maximum expected write concurrency without overflow; if the buffer approaches capacity, back-pressure is applied to the write path so that the flush rate keeps pace with the ingestion rate.
Sequential reads on SMR drives are unaffected by the write geometry. The read head can address any track directly and data is read in the order it was written, which is sequential within each band. Read throughput on a well-managed SMR drive is comparable to CMR drives at the same platter density.
Why This Matters More Than It Sounds
The software framing matters here: SMR's write constraint is not a performance bug to be worked around, it is a geometric property of the recording medium. Drives that expose this constraint directly (HM-SMR) require host software that understands it. Drives that hide it through cache buffering (DM-SMR) trade predictable throughput for drop-in compatibility, but at the cost of periodic throughput collapses when the cache drains.
For AI training data storage, the workload pattern (large sequential writes, large sequential reads) aligns well with SMR's capabilities when managed correctly. The areal density advantage translates directly into lower cost per terabyte at equivalent physical drive size. Understanding the geometry is the prerequisite for deploying SMR drives correctly, which is why we wrote this before the benchmarks. The numbers only make sense once the underlying mechanism is clear.