Garbage collection optimization with data separation for large data storage in deep learning applications

Abstract

Deep learning has revolutionized numerous domains, creating an urgent need for storage systems capable of handling massive datasets and the intensive computational demands inherent to these workloads. Solid-State Drives (SSDs), known for their fast random access, low power consumption, and shock resistance, have emerged as a preferred storage medium in this context. However, traditional SSDs face critical challenges, including garbage collection (GC) overhead, write amplification, and inefficiencies in the software storage stack, stemming from the intrinsic characteristics of NAND flash and limitations in the existing storage ecosystem. These challenges underscore the necessity for specialized SSD controller chip designs tailored for deep learning workloads, addressing performance bottlenecks and optimizing data management to meet the unique demands of AI-driven applications. In this work, we implemented an open-channel SSD (OCSSD) based on a Xilinx FPGA, which can effectively alleviate the above-mentioned issues by exposing the structural characteristics of NAND flash to the host. To mitigate the performance cliff of I/O requests during GC operations, the link distance for data transmission is shortened by decoupling the host end and the device end. Moreover, the valid data migration and the GC operation frequency are both dramatically reduced by detecting and separating hot data and cold data to improve the overall performance of the SSD system. To verify the superiority of our design, we build a test platform through hardware and software co-design. The experimental results show that random read and random write bandwidth are increased by 159.7 % and 25.3 % compared to the mainstream SSDs, respectively. The latency of a single GC operation is reduced by an average of 12.64 % and the GC frequency is lowered by up to 64.8 %.