Narayanan Sundaram, Research Scientist, Intel Labs
summary
GraphMat implements a vertex-centric front end but compiles those primitives to generalized sparse matrix-vector and matrix-matrix operations, yielding a backend tuned with HPC techniques. Benchmarks show it outperforms GraphLab/Giraph/PowerGraph by orders of magnitude in distributed scenarios and runs algorithms within 20% of hand-optimized native C for several graph kernels. Single-node code (handles ~0.5B edges on 64 GB RAM) is BSD open source on GitHub and a Spark integration is available to replace other frameworks’ graph operators.
AI-written summary · about
chapters (8)
- 0:00GraphMat and productivity gap
- 3:32Graph analytics needs faster frameworks
- 7:53GraphMat’s vertex programming backend
- 13:15Transforming vertex programs to matrix ops
- 18:13Performance vs other frameworks
- 20:43Scalability and real-world gains
- 22:34Summary and future plans
- 23:19Q&A: Memory limits and distributed code
related talks
29mTed Willke, Senior Principal Engineer & GM, Datacenter Group, Intel
57mGraphBLAS: Building a C++ Matrix API for Graph Algorithms
12mMake ML outcomes roar with graph neural networks on TigerGraph and AWS (PRT097)
28mNetworkX is Fast Now: Graph Analytics Unleashed
32mTowards new deep learning abstractions on top of exist frameworks
56mGraph Analytic Systems
44mScalable Algorithmic Primitives for Data Science
42mAnima Anadkumar, Principal Scientist, Amazon Web Services, Endowed Professor, CalTech