PASM: Parallelism Aware Space Management strategy for hybrid SSD towards in-storage DNN training acceleration

Abstract

With the explosive growth of data volume and great improvement in flash technologies, SSD-based In-Storage Computing (ISC) is becoming one of the most important means to accelerate data-intensive applications, which also provides a possibility for in-storage acceleration of DNN training. However, the continuous write requests involved in DNN training become a big challenge for the reliability and efficiency of the flash-centric computing system. Recently trend in high-density and low-cost NAND flash further aggravates this challenge. To address this problem, we propose a Parallelism Aware Space Management (PASM) strategy to enable the utilization of SLC-TLC hybrid SSD for DNN training acceleration. And two key technologies are included in PASM. Firstly, we proposed the conflict-free parallelized data layout. It distributes data of different lifetimes into different kinds of physical blocks to avoid data migration as far as possible. Meanwhile, it enables parallel read/program and erase operations by exploiting the inherent parallelism in the flash array. As a result, it eliminates the conflicts between continuous I/O requests and garbage collections so as to provide a stable I/O performance during training. Secondly, a novel garbage collection strategy, called lifetime-aware deterministic garbage collection, is proposed. This scheme performs erase operations without disturbing continuous I/O requests by sensing the data lifetime and tightly coordinating the neural network training process, which is able to reclaim invalid blocks in a timely and efficient way to guarantee that the high endurance of SLC can be utilized continuously. Finally, to verify the performance of PASM, we compared PASM with the related state-of-the-art hybrid SSD and TLC-only SSD under Resnet50 training workload. Experimental results show that the PASM improves I/O performance by 20% and increases lifetime by 6.6 times.