In large-scale high-performance computing (HPC) systems, various applications with different I/O characteristics are executed. However, underlying HPC storage systems, such as a parallel file system (PFS) and a global and node-local burst buffers (BBs...
In large-scale high-performance computing (HPC) systems, various applications with different I/O characteristics are executed. However, underlying HPC storage systems, such as a parallel file system (PFS) and a global and node-local burst buffers (BBs), are unaware of these characteristics of applications, and using a single fixed file layout for all applications can degrade the performance. Moreover, utilizing faster storage tiers (i.e., BB) with high-performing storage capabilities can introduce some challenges, such as manual configurations, explicit data management, and unexpected I/O performance degradation, which result in low utilization rates and inefficient storage utilization.
In this dissertation, we analyzed data access patterns of applications in HPC systems such as Perlmutter at NERSC and Nurion at KISTI. Specifically, we divide data access patterns into spatial and temporal file access patterns. Based on analyses, we autonomously configure the HPC storage systems and optimize the I/O performance of applications by (1) adjusting file layout adaptively in real-time and (2) redirecting and relocating file accesses within different tiers in the storage hierarchy of the HPC system.
To do this, we propose schemes to improve the I/O performance of applications by leveraging spatial and temporal file access patterns. First, we design A2FL, an autonomous and adaptive file layout adjustment scheme that optimizes PFS configurations by analyzing the spatial file access pattern of the applications. A2FL intercepts and records spatial file access patterns from I/O operations of applications in real-time, transforms them into graphical representations that are used for predicting I/O performance, and adjusts file layouts adaptively. Second, we propose AutoR2, an autonomous I/O redirection and file relocation scheme for HPC storage systems. The goal of AutoR2 is to autonomously redirect I/O applications to a faster storage tier based on application I/O and system performance characteristics. AutoR2 analyzes I/O characteristics and performance trends on different storage tiers. Then, it learns temporal file access patterns to manage file data among HPC storage tiers without user intervention.
To demonstrate the effectiveness of the proposed schemes, we evaluate the schemes on production HPC systems. Our evaluation results show that A2FL significantly enhances I/O performance, with improvements of up to 68.35x compared to the default file layout. Moreover, AutoR2 can improve by up to 1.63x speed-up in terms of accumulated I/O time compared with the default system.