Thesis
PMIO AND SCISSD: STORAGE ACROSS MEDIUMS AND PLATFORMS
Washington State University
Master of Science (MS), Washington State University
05/2024
DOI:
https://doi.org/10.7273/000006968
Abstract
Storage is a fundamental part of modern computation. New storage platforms are being explored, and it is important to analyze their effectiveness across many domains. In this work, we explore two contemporary storage platforms, computational storage and persistent memory. We analyze how they perform in two separate projects.
First, we analyze the performance of persistent memory as a replacement for Collective I/O buffers for high-performance computing (HPC) applications. Collective I/O is a portion of the MPI-IO layer designed to reduce the number of small requests issued by processes running in HPC environments. During Collective I/O, data is shuffled between processes and temporarily stored within DRAM buffers to form more contiguous requests. Current implementations suffer from high amounts of inter-process communication, as well as limitations on performance due to the limited size of the DRAM buffer. PMIO seeks to replace the DRAM buffer with a buffer in persistent memory. In addition, we introduce a novel log-based buffer to fully exploit the sequential performance of persistent memory. Finally, we develop a merging algorithm to reduce inter-process communication. Analysis shows a performance increase of up to 151X and 121X for writes and reads respectively on the Perlmutter supercomputer.
Following this, we explore offloading augmentation in the pipeline of machine-learning training to Computational Storage Devices (CSD). In modern machine learning applications, samples within a training loop are often "augmented" as they are used for training. During this process, a series of filters are applied to the samples to improve diversity. The augmentation process can be costly, as it forms a bottleneck before data can be sent to GPUs. We seek to remove this bottleneck by offloading the augmentation processed to CSDs. In doing so, we will augment the samples asynchronously such that whenever an epoch finishes, a new set of pre-augmented samples will be available. To this end, we develop SciSSD, a set of API's, applications, and CSD firmware to enable the offloading and management of data on CSDs. In the best case, SciSSD can reduce reading times by more than a minute per epoch.
Metrics
3 File views/ downloads
10 Record Views
Details
- Title
- PMIO AND SCISSD
- Creators
- Keegan Isaiah Honor Sanchez
- Contributors
- Xuechen Zhang (Chair)Ben McCamish (Committee Member)Xinghui Zhao (Committee Member)
- Awarding Institution
- Washington State University
- Academic Unit
- School of Engineering and Computer Science (VANC)
- Theses and Dissertations
- Master of Science (MS), Washington State University
- Publisher
- Washington State University
- Number of pages
- 57
- Identifiers
- 99901125140001842
- Language
- English
- Resource Type
- Thesis