After four years of SMR storage, here's what we love—and what comes next

Four years of SMR at Dropbox: density, power, and cost lessons

Dropbox began deploying shingled magnetic recording (SMR) drives in 2018, becoming the first major tech company to adopt the technology at scale. By 2019, roughly 25% of the fleet was SMR-capable. Today, 90% of total HDD capacity is SMR, spread across three drive models.

The shift has delivered measurable gains in storage density, power efficiency, and cost per exabyte. It has also reshaped how Dropbox evaluates hardware and collaborates with vendors. Here's a look at what the past four years have shown.

Density gains from recording techniques and enclosures

SMR's core advantage comes from shingling: data tracks overlap like roof shingles, boosting tracks per inch (TPI) by roughly 20% over conventional perpendicular magnetic recording (PMR). Track width is no longer constrained by the write head but by the narrower read head.

Newer SMR models add two-dimensional magnetic recording (TDMR), which pairs an extra reader per head. The second reader improves signal detection and cancels inter-track interference noise, allowing even narrower tracks and higher areal density.

Platter technology has also advanced. Thinner platters and higher counts per enclosure mean more surface area for bits. Recent drives pack nine platters and eighteen heads—more than double the first drives Dropbox deployed. Platter spacing is now measured in nanometers.

Enclosure design has evolved in parallel. Dropbox has increased the number of 3.5" HDDs per server by 2.8x since the first-generation design and 1.8x since SMR adoption began. The tradeoff is thermal management: cooling must stay within the 45°C HDD window. Future gains will likely come from thermal and vibration efficiency rather than raw enclosure count.

Power per terabyte keeps falling

Power consumption per terabyte (TB/watt) has dropped 5-6x since the first 4 TB deployment. SMR packs more capacity into the same physical and energy footprint as PMR.

The first 14 TB SMR drive nearly halved power consumption for idle and random read workloads versus its PMR predecessor. Current 18 TB and 20 TB drives run at roughly .30 watts per TB idle and .50 watts per TB for random reads. Vendor projections indicate this trend continues at higher capacities.

At the exabyte scale, the savings compound. The sixth-generation storage server requires five fewer megawatts and a third of the physical rack space per exabyte compared to the fourth-generation design. Lower power draw is a key part of Dropbox's data center sustainability efforts.

Cost per exabyte and footprint shrink

With SMR at 90% of fleet capacity, the cost benefits first identified in year one remain intact. SMR drives store 10-20% more data than equivalent PMR drives at little to no cost premium. Over time, that translates into fewer drives for the same capacity, substantially reducing $/exabyte costs.

An unexpected outcome: total physical footprint has decreased even as capacity grew. Smaller 4 TB, 6 TB, and 8 TB drives were decommissioned and replaced by higher-density models—a single SMR drive can replace four or five 4 TB units.

A smaller fleet means fewer repairs, less management overhead, and reduced space and power requirements in data centers. These savings have compounded over time.

Software ecosystem matures around zoned storage

The software stack supporting SMR has matured significantly. Initial Dropbox deployments used a custom disk format with the libzbc library for read/write operations, bypassing any filesystem.

Current efforts incorporate zoned device support in the Linux kernel and libraries like libzbd. These eliminate the need for custom disk formats. libzbd offers asynchronous I/O, higher queue depth, priority-based I/O, block trace tool compatibility, and lower overhead than libzbc.

Dropbox's internal SMRtest tool, which generates synthetic production-like workloads, has been updated with libzbd support, multi-threading, and 4k sector sizes. Zoned block device support also landed in fio with version 3.9. A mix of these test suites is used with ecosystem partners and vendors depending on the use case.

Closer vendor collaboration catches bugs earlier

Operating one of the largest host-managed SMR fleets has deepened the evaluation relationship with HDD partners. The key change: vendors now run Dropbox workloads at scale on identical storage hardware at their own sites. Dropbox also built an in-house simulator of Magic Pocket to gain production-like signal earlier in hardware evaluation.

This approach improves the testing pipeline in several ways:

• Reproducibility: identical hardware, software, and firmware stacks on both sides eliminate interoperability issues that don't apply to production.
• Velocity: hardware physically close to experts shortens root-cause analysis. An engineer can pull a SATA trace, develop a fix, test a patch, and iterate without coordinating across time zones.
• Scale: some bugs only surface after significant drive hours. At-scale testing deploys 5,000+ drives, often exceeding what vendors test on their own.
• Early signal: the Magic Pocket simulator catches issues before new drive models reach production.

These collaborations have enabled two new SMR deployments that have accumulated several billion production drive hours. Feedback shared with vendors regularly finds its way into publicly available hardware and software, benefiting the wider SMR ecosystem.

What Comes After SMR

The areal density gains that SMR and two-dimensional magnetic recording (TDMR) have delivered over the past several years are beginning to plateau. Further incremental progress remains possible—more platters, modest density improvements—but the next major jump requires a fundamentally different recording technology. That jump is heat-assisted magnetic recording (HAMR).

HAMR addresses what researchers call the magnetic recording trilemma. Higher densities demand smaller media grains, but smaller grains need higher magnetic anisotropy to remain thermally stable, which in turn demands a stronger head field to switch them. HAMR sidesteps that conflict by using the head to spot-heat the media to roughly 450°C during the write operation. The heating, which happens in nanoseconds, temporarily softens a new disk medium made of iron-platinum (FePT) material. That allows the media to support smaller, thermally stable grains and smaller bits, yielding higher areal density than SMR's overlapping-track approach. Dropbox expects HAMR-based drives to open a path to capacities of 50 TB and beyond.

Adopting HAMR means new read/write head designs, new media, updated firmware, and a different manufacturing process. But from the host's perspective, HAMR drives should appear and behave like conventional HDDs. Dropbox also points to partner data indicating reliability and manufacturability on par with SMR.

A New Sensitivity to Vibration

With higher-density drives on the horizon, Dropbox's design focus has shifted away from simply packing more disks into an enclosure. As data tracks get smaller and closer together, the margin for physical vibration shrinks. High-frequency vibration—from fans, from the seek activity and rotational forces of neighboring drives, even from a single drive's own operation—can cause head positioning errors that degrade I/O performance. The tolerances that older, lower-density designs could absorb are no longer sufficient.

Future work will emphasize suppressing structural vibration in the chassis and reducing fan noise to limit performance degradation. Dropbox is also planning to adopt the Open Compute Project's HDD Acoustical Surrogate specification, a standardized vibrational testing approach, in its seventh-generation designs.

Rethinking Infrastructure Around Density

Higher-density drives expose bottlenecks beyond the mechanics of the drive itself. Dropbox's seventh-generation server is designed to support more than 6 PB in a single enclosure, which required re-architecting the network so data can be drained and repaired at an acceptable rate. The company also anticipates that dual-actuator drives may eventually be necessary to meet I/O-per-terabyte requirements. A single I/O channel caps the sustained I/O a drive can deliver, but a second channel would effectively double the I/O operations per second (IOPS) available from each drive.

Looking further out, the combination of nanometer-scale track widths, recording heads flying roughly a nanometer above the disk surface, new platter materials, and HAMR provides a roadmap toward 100 TB HDDs. After four years of production SMR experience, Dropbox sees those developments as the next phase of storage density—one it expects to be just as substantial as the transition to SMR was.