Taurus system accelerates out-of-core GNN inference on massive graphs

Researchers have developed Taurus, a novel system designed to accelerate out-of-core Graph Neural Network (GNN) inference on massive graphs that exceed available RAM. Taurus optimizes inference by reformulating layer-wise computations as sequential SSD scans, utilizing a pipelined hierarchy of GPUs, CPUs, and SSDs. This approach significantly reduces I/O costs and communication times compared to existing distributed and disk-based GNN systems. In benchmarks on graphs with up to 269 million vertices and 514 GiB of features, Taurus demonstrated performance improvements of 7-25x over layer-wise baselines and 40-140x over vertex-wise baselines. AI

IMPACT This system could enable more efficient AI model inference on extremely large graph datasets, potentially impacting fields that rely on graph-based AI.

RANK_REASON The item describes a new system and benchmark results presented in an arXiv paper. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• arXiv
• central processing unit
• graphics processing unit
• graph neural network
• Taurus

• infra
• paper

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →