Did you know that the vast majority of data in the world, especially with hyper scalers, is still on hard drives? While solid-state media certainly dominates the news, you can’t discount the fact that magnetic storage in the form of hard disk drives (HDDs) is still the dominant storage media. Not only that, but hard drives are continuing to evolve to meet the demands of today’s workloads. HDD’s continue to evolve with energy assisted recording to increase the density and multiple actuators to increase the performance. However, hard drive interfaces have been largely the same SAS or SATA that have been around for 15+ years, with mostly interface speed bumps over the years. Should we re-think the interface for the next decade of HDDs?
Most storage platforms today leverage SATA or SAS hard drives that sit behind an I/O controller that is either discrete or embedded into a chipset. What if we could eliminate the extra complexity and the legacy proprietary nature of most IO controllers/drivers and connect the HDDs directly to the CPU? What if we could address the data on magnetic storage directly like we do on today’s NVMe SSDs? Well, We can with NVMe hard drives!
The SSD market moved to NVMe to improve performance and reduce latency. People often ask if this is required or even helpful for hard drives, since magnetic storage is an order of magnitude slower IOPS and much higher latency than NAND. The short answer is no from a per-drive link speed. SAS and SATA offer plenty of throughput today. That said, the main value of moving HDDs to NVMe is to simplify the topologies and remove the extra translation layers of SATA and SAS. Furthermore, for dual and future quad actuator drives, the concept of NVMe namespaces is a logical way to address each actuator. This is where unifying the interfaces between SSDs and HDDs can simplify storage stacks.
In the Open Compute Storage workgroup, we started an NVMe HDD workstream in late 2020 to help explore and define the high level requirements of what we need from an NMVe HDD. The primary focus has been to review the formfactor, connector, power, and signal requirements. Once we reach consensus and publish a draft specification, we will share this with the standards organizations outside of OCP to formalize the details and drive and industry standard.
The NVMe workgroup has already started working on the basic interface requirements to enable an NVMe HDD.
Please join us to define the next decade of HDD interfaces!