Enhancing SSD Performance for GPUDirect Storage Systems Through Dynamic Address Allocation and Fine-Grained Address Mapping
Author(s)
Arora, Sachitt
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
The recent boom in machine learning (ML) and high performance computing (HPC) has necessitated efficient data processing to keep up with the increased rate of innovation. New machine learning and scientific computing models require larger datasets year-over-year, and this motivates the optimization of the data processing path related to a system’s storage. Furthermore, research has been increasingly using GPUDirectStorage systems to eliminate the CPU as an unnecessary middleman in the data processing involved in several applications. In this research, we examine several optimizations related to improving SSD performance within these GPUDirectStorage systems including fine-grained address mapping and token-based garbage collection. From our results, we found that these changes, alongside other optimizations, resulted in higher Input/Output Operations per Second (IOPS) and minimized tail latencies, overall increasing the efficiency of this type of architecture. Fine-grained address mapping enables write operations being done at a smaller granularity, reducing redundancy and unnecessary data invalidation. Dynamic address allocation better exploits parallelism within the SSD by assigning I/O operations to different planes based on current resource usage. Our findings suggest that optimizing the SSD configuration in this way for GPUDirectStorage systems can significantly speed up the processing of large applications, which can increase innovation turnaround in fields such as healthcare, finance, and language. Furthermore, the reduced data movement and enhanced resource utilization can result in longer SSD lifespans and less energy consumption, providing a more cost-effective solution to meet data processing demands.
Sponsor
Date
Extent
Resource Type
Text
Resource Subtype
Undergraduate Research Option Thesis