Research on Frontier Technologies in Data Center and Server
The rapid growths of data volume and data-intensive applications call for large-capacity, low-cost, and high-performance storage devices. The standard block interface has long been used and performed well in conventional storage devices, i.e., hard disk drives (HDDs). It provides an easy-to-use abstraction, i.e., a linear array of logical block addresses (LBAs) that can be read and written randomly, conforming to the media access characteristics of HDDs.
However, the block interface has increasingly become an obstacle for modern flash-based solid-state drives (SSDs). Due to higher performance, reliability, and energy efficiency, SSDs have been replacing HDDs and widely deployed in various storage systems. Flash memory and must be erased in bulk before being overwritten and only allows sequential writes in an erase unit. To address the mismatch between such features and the block interface abstraction, an SSD employs a complicated firmware, called flash translation layer (FTL). Not only the hardware cost increases as significant amounts of computing, memory, and over-provisioned flash resources are required, but also software overheads are introduced and degrade the SSD performance and lifetime.
Recently, the zoned namespace (ZNS) interface, standardized by NVMe, is newly proposed for SSDs. The ZNS interface abstracts a storage device as an array of logical zones. Each zone contains contiguous LBAs that can only be written sequentially. This abstraction matches the write features of flash memory and thus can substantially simplify the FTL and reduce the hardware cost and software overheads. Furthermore, the ZNS interface promotes hardware and software co-optimizations by allowing host software to make application-aware data placement and I/O scheduling on flash memory. Overall, ZNS SSDs have emerged as a promising solution for building high-capacity, low-cost, and high-performance storage systems. How to design ZNS SSDs and apply them to current systems has become an interesting and hot research topic in both academia and industry.
One critical problem, which hinders the adoption and potentials of ZNS SSDs, originates from the fixed zone size restriction. For data reliability, the FTL usually organizes a line of flash blocks across many parallel dies/chips into a superblock or RAID stripe with parity. A superblock also becomes the unit of flash erasures and space allocation for new writes. To match the zone abstraction and flash layout, a ZNS SSD typically maps each zone to a dedicated superblock, whose size is fixed and large (e.g., a few gigabytes).
However, the hardware limit of a fixed and large zone size results in several drawbacks. Especially, ZNS SSDs from different manufactures or based on different flash models may have distinct zone sizes. First, a fixed zone size would prevent applications (which may differ in favorable zone sizes) from optimizing data placement according to their unique needs. Second, a large zone size would cause high garbage collection overheads, since host software has to relocate massive data between zones. Third, it is challenging to implement some advanced features based on ZNS SSDs, such as constructing a RAID system over ZNS SSDs with different zone sizes, performing fine-grained data management on a ZNS SSD with a large zone size.
- A comprehensive analysis about the impacts of zone sizes of ZNS SSDs on the data management and performance of applications
- Effective approaches to address the zone size limitations of ZNS SSDs and the challenges of applying ZNS SSDs to current storage systems
- Publications of 1-2 papers in top-tier academic conferences and journals
Related Research Topics
- Flash translation layer of ZNS SSDs
- Data management based on ZNS SSDs
- Hardware and software co-designs of zoned storage systems