GPU-Initiated On-Demand High-Throughput Storage Access in the BaM System Architecture
- 1. UIUC
- 2. NVIDIA
- 3. University at Buffalo
- 4. IBM Research
- 5. NVIDIA/Stanford
- 6. NVIDIA/UIUC
Description
BaM proposes a novel system architecture that offers mechanisms for GPU code to efficiently access storage and enable GPU self-orchestrated storage access. As such, with the artifact, we want to establish the following key goals:
(1) BaM is functional with a single SSD and supports the contributions described in the paper,
(2) BaM is functional with multiple SSDs - up to 2 SSD can be tested on the provided system,
(3) BaM is functional with different types of SSDs - consumer-grade Samsung 980 Pro and datacenter-grade Intel Optane SSDs are available in the provided system,
(4) BaM can work with different applications - microbenchmarks and graph applications (w/ real datasets) are provided for ASPLOS'23 AOE.
Notes
Files
Files
(3.2 MB)
| Name | Size | |
|---|---|---|
|
md5:4daa22208a0d9540e6c79830567f7203
|
3.2 MB | Download |
Additional details
Related works
- Has part
- Software: https://github.com/ZaidQureshi/bam.git (URL)
- Is described by
- Conference paper: 10.1145/3575693.3575748 (DOI)
- Is published in
- Preprint: https://arxiv.org/abs/2203.04910 (URL)