GLIST: Towards In-Storage Graph Learning
summary
GLIST moves GNN computation into SSDs using an in-storage FPGA accelerator plus an embedded CPU, flash arrays and DRAM, implements sampler / aggregate / combine units and offers graph update/registration/model-registration/analysis APIs. It reorganizes property vectors on 16 KB flash pages and reorders requests to exploit spatial locality and temporal locality, reducing data movement. An FPGA prototype on the COSMOS OpenSSD platform achieves about 13.2× and 10.1× speedup over CPU and GPU baselines and cuts power use by up to 98.7% while handling node/link/graph benchmarks for datasets >100 GB.
AI-written summary · about
chapters (16)
- 0:00Intro
- 0:24Background of Graph Learning
- 1:50Background of GNN-based GL
- 2:31Obtain Embedding Vectors with GNN - 2
- 3:27Predicting Relationship
- 3:51GNN-based GL Workload Characterization
- 5:01Challenges and Solutions 10 bottleneck and memory constraints
- 5:43Outline
- 6:13System Overview
- 8:09GLIST User Library - Graph Reorganization
- 9:55In-Storage Graph Learning Accelerator
- 10:25Experimental Setup Hardware
- 11:17Evaluation - GLIST Performance
- 11:37Evaluation - GLIST Efficiency
- 11:57Evaluation - GLIST Optimization
- 12:47Conclusion
related talks
13mHardware/Software Co-Programmable Framework for Computational SSDs to Accelerate Deep...
16mCognitive SSD: A Deep Learning Engine for In-Storage Data Retrieval
17mBGL: GPU-Efficient GNN Training by Optimizing Graph Data I/O and Preprocessing
16mAccelerating Distributed Graph Learning by Using Collaborative In-Network...
16mPreview: Graph NNs
19mOptimizations and abstractions for sparse machine learning
15mMGG: Accelerating Graph Neural Networks with Fine-Grained Intra-Kernel Communication...
23mSparse Computing Drives Energy-Efficient Computation for Artificial Intelligence