HDDs and SSDs. What are the right questions?

Technology replacement is a familiar tale — the passing of the torch from old to new. As an example, steam engines displaced by internal combustion engines, bulky tube displays bowed out gracefully and were replaced with sleek LCD screens. One very close to my heart was the near-complete replacement of traditional hard disk drives (HDDs) with solid-state drives (SSDs) in laptops.

With over three decades of involvement in flash memory and flash-based SSD architecture and development, I am well attuned to storage technology shifts. I’ve witnessed pivotal transitions in SSDs — the transition from NOR and NAND, shifting from single bit per cell (SLC) to multiple bits per cell (MLC, TLC, PLC, etc.), transitioning from SATA to PCIe/NVMe, and the shift from 2.5” HDD form factors to dedicated SSD form factors such as M.2 and EDSFF. Each has significantly impacted the industry overall and propelled us forward.

One of the prominent questions that consistently arises is: When will SSDs become more cost-effective than HDDs and fully replace them? According to IDC¹, in 2023, about 13% of the PB shipped to data centers were flash SSDs (predominantly TLC) compared to capacity-optimized HDDs (predominantly 20TB 3.5”). With SSD dollar per terabyte decline at around 17% CAGR and capacity-optimized HDD at 8.5% CAGR, you could forecast a theoretical cross over 10-15 years from now.

In this blog, I’ll explore what might be a better question to ask and offer a prognostication of what the future holds for data storage. I am building upon a recent Micron blog by Currie Munce² and discussing a single hyperscaler’s experience transitioning from a strictly HDD based warm tier file system and the impacts AI forced upon them and the solution, which included SSDs as a tiered caching layer.

Meta’s warm tier transition to a composite SSD/HDD solution driven by AI

In 2021, Meta presented the Tectonic file system at the Usenix File and Storage Technology conference³. This innovative system combines low-latency, small-sized blob storage with quick required provisioning of HDDs prioritizing IOPS with data warehousing, which could provision HDDs prioritizing density. Tectonic operates as a unified exabyte-scale system built using thousands of storage nodes comprising 72 3.5” HDDs per shelf. The storage nodes are based on 3.5” capacity-optimized HDDs.

All was good until the storage demands of AI hit hard, showing that within about a year, their online ingestion bandwidth increased four times! This exceeded the 100% HDD solution’s peak I/O demand.

To continue with an HDD-only warm tier would require significant over-provisioning of HDDs for I/O, resulting in excess captive storage capacity with high cost and power consumption. They considered this against the option of either replacing the warm tier with 100% SSDs or adopting a composite solution of HDDs and SSDs. The table below, published by Meta, illustrates the challenges: significant over-provisioning with HDDs alone, the difficulty of replacing with SSDs only at current densities, and a composite HDD and flash storage solution that strikes an ideal balance.

Architecting the right storage caching

The revised warm tier solution was named Tectonic-Shift. It included an application transparent TLC cache supporting the HDDs already in the warm tier. There is a great discussion in the paper presented at ACM ’23⁴ about how the authors selected their caching policies after a detailed analysis of AI traces and their unique properties and determining the right tradeoffs between insertion and eviction policies trading off performance against power, cost and endurance of the SSDs in the warm tier cache.

The resulting insertion of an SSD caching layer into Meta’s HDD-based Tectonic-Shift system given the explosion of AI workloads has absorbed the increased workload.

Composite SSD+HDD beyond AI. Looking forward.

A critical tradeoff lies not only in achieving the right storage density but also in ensuring the appropriate I/O performance for that density. As previously discussed in Currie Munce’s blog, a useful metric to consider is performance divided by density. Let’s illustrate this with an example: In the case of “Facebook’s Tectonic Filesystem: Efficiency from Exascale” paper, the authors proposed a composite cluster requiring 100 petabytes (PB) at a peak rate of 10 terabytes per second (TB/s). This translates to a storage throughput density of approximately 100 megabytes per second per terabyte (MB/s/TB) at peak performance. However, the average requirements differ across workloads. For AI workloads, the suggested storage density average is around 20 MB/s/TB, while object stores typically operate at approximately 5 MB/s/TB. Blob storage, on the other hand, hovers around 2 MB/s/TB²

Balancing performance and density becomes crucial as we navigate the evolving landscape of data storage.

Micron 6500 ION purpose-built for the warm tier

Although HDDs have had a historically impressive density CAGR, their performance has been near flat. Consequently, the megabytes per second per terabyte (MB/s/TB) decreases with each successive generation. Micron believes this will be addressed with further SSD tiering versus HDD over provisioning. SSDs will displace HDDs in the warm tier. HDDs will continue to serve the cooler/cold tiers.

So the right question is:

“When will SSDs displace (not replace) HDDs meaningfully in the data center?”

And the answer is:

They already have!”

This is precisely why Micron developed and launched the Micron 6500 ION SSD. This award-winning SSD optimizes TCO in tiered storage solutions through high throughput density, power efficiency, and low latency.

This becomes the complementary tier to HDDs and not the replacement of HDDs for the foreseeable future.