Scaph: Scalable GPU-Accelerated Graph Processing with Value-Driven Differential...
summary
Scaph is a GPU-accelerated system that partitions graphs into host-resident subgraphs and adaptively classifies each iteration whether a subgraph is high-value or low-value, using value-driven differential scheduling and two engines that either stream whole subgraphs or transfer only neighbor data for active vertices. It implements bitmap-based UD extraction, chunked GPU tiles and a re-scheduling “queen” for GPU-resident subgraphs, uses thresholds (α≈50%, β≈30%), and reports mean speedups versus Totem 4.12×, Graphie 8.93× and Garaph 3.71× on real and synthetic large graphs.
AI-written summary · about
chapters (6)
- 0:00Start
- 0:38Graph Processing: CPU vs. GPU GPU V100
- 3:34A Generic Example of Graph Processing Engine
- 6:13Characterization of Subgraph Data
- 10:26Value-Driven Differential Scheduling
- 14:09Low-Value Subgraph Processing • NUMA Aware Load Balancing
related talks
12mTEGRA: Efficient Ad-Hoc Analytics on Evolving Graphs
24mLarge Scale Graph Computation with Apache Giraph
30mScalable Graph-Parallel Inference Algorithms and Systems
15mMGG: Accelerating Graph Neural Networks with Fine-Grained Intra-Kernel Communication...
45mScale Learning and Reasoning Across Heterogeneous Gradients and Semantics
39mThorny path to the Large Scale Graph Processing
28mOptimizing Large-Scale Model Training with Ray Compiled Graphs
18mGPU-Disaggregated Serving for Deep Learning Recommendation Models at Scale