Kioxia RAID Offload: Faster NVMe Performance
The Future of RAID: Kioxia’s SSD Offload Technology and the Evolution of Data Storage
Table of Contents
As of July 8,2024,the data storage landscape is undergoing a quiet revolution. While much attention is focused on the relentless pursuit of faster processors and more memory, a critical bottleneck has been looming: the limitations of customary RAID configurations in the face of increasingly powerful solid State Drives (SSDs). Kioxia,a leading innovator in memory technology,recently demonstrated a proof-of-concept solution at Flash Memory Summit (FMS) 2024 that promises to redefine how RAID operates,perhaps unlocking critically important performance gains and reducing system overhead. This article delves into Kioxia’s innovative SSD offload methodology, its implications for enterprise storage, and its potential to become a new standard in the industry.
The RAID Bottleneck in the Age of fast ssds
Redundant Array of Independent Disks (RAID) has long been a cornerstone of data storage, providing both performance enhancements and data redundancy. however, the fundamental principles of RAID, designed for the mechanical limitations of Hard Disk Drives (HDDs), are increasingly at odds with the speed and characteristics of modern SSDs.
Traditional RAID levels, such as RAID 5 and RAID 6, rely on parity calculations to ensure data integrity. In a typical write operation with RAID 5, for example, the system must read data from multiple drives, calculate the parity, and then write both the data and the parity data to different drives. This process, even with dedicated RAID hardware controllers, introduces significant overhead.
The problem is exacerbated by the sheer speed of SSDs. As SSDs become faster with each generation,the RAID controller – or even the CPU handling software RAID – struggles to keep pace. The data has to travel back to the CPU and main memory for processing before being written, creating a bottleneck that negates the benefits of the fast storage. This is particularly true in scenarios without hardware acceleration, where the CPU bears the entire burden of RAID calculations. The result is diminishing returns on SSD performance, and increased CPU and DRAM utilization.
Kioxia’s SSD Offload Methodology: A Paradigm Shift
Kioxia’s proposed solution,demonstrated at FMS 2024,tackles this bottleneck head-on by shifting the RAID processing burden directly onto the SSDs themselves. This is achieved through a combination of PCIe Direct Memory access (DMA) and the utilization of the SSD controller’s Controller Memory Buffer (CMB).
Understanding the Key Components:
PCIe DMA: pcie DMA allows devices to access system memory directly, bypassing the CPU. This is crucial for high-speed data transfer. Controller Memory Buffer (CMB): The CMB is a dedicated memory area within the SSD controller used for temporary data storage and processing.
On-SSD Parity Accelerator: Kioxia has integrated a dedicated accelerator block within the SSD controller specifically designed for parity computation.
How it Works:
Instead of sending data to the CPU for RAID calculations, Kioxia’s offload scheme leverages PCIe DMA to allow the SSD controller to directly access the CMB of neighboring SSDs within the RAID array.The DMA engine can access the entire host address space,including the BAR-mapped CMB of peer ssds. This enables the SSDs to communicate and exchange data without CPU intervention.
The on-SSD parity accelerator then performs the necessary parity calculations within the SSD controller itself. The resulting parity data is written directly to the appropriate SSD, completing the RAID operation with minimal CPU involvement. This drastically reduces latency and frees up valuable CPU resources.
Performance Gains and System Impact
The results of Kioxia’s proof-of-concept implementation are compelling. according to Kioxia, the offload scheme achieved:
Nearly 50% Reduction in CPU Utilization: By offloading the RAID calculations to the SSDs, the CPU is freed up to handle other tasks, improving overall system performance.
Over 90% Reduction in System DRAM Utilization: The reduced need to buffer data in system memory substantially lowers DRAM usage, further optimizing system resources.
Efficient Scrubbing Operations: The on-SSD parity accelerator can also handle background scrubbing operations (data integrity checks) without impacting CPU performance.
These gains are particularly significant for data-intensive applications such as databases, virtualization, and high-performance computing, where every cycle of CPU and memory bandwidth counts. The reduction in latency also translates to faster response times and improved user experience.
Standardization and the NVM Express Working Group
Kioxia recognizes that the full potential of this technology can only be realized through widespread adoption. To that end, they have already begun the process of