High-Performance Logic Programming on Servers, Clusters, and GPUs
summary
A new suite of engines targets Datalog on servers, clusters, and GPUs with concrete speed and scale claims: Ascent is a Rust-embedded engine that integrates custom relation providers and scales well to ~32 cores (used to build a 25k-line LLVM analysis and enumerate billions of control-flow points); Slog runs an MPI runtime and scales to roughly 16–32k threads on Theta; GDlog (CUDA/HIP) uses a HISSa index and reports up to ~200× speedups versus CPU (H100 vs AMD Epic), e.g. 14s vs 232s on a large transitive-closure workload.
AI-written summary · about
chapters (12)
- 0:00Workshop introduction
- 7:24What is Datalog
- 13:50Datalog systems overview
- 16:41Ascent in Rust
- 21:56Compiling Datalog efficiently
- 29:04Semi-naive evaluation
- 33:48Custom data structures in Ascent
- 37:46Scaling Datalog on clusters
- 42:00GPU acceleration for Datalog
- 44:07GDLog engine design
- 47:37Performance on large graphs
- 50:46Open-source tools
related talks
22mOatlog: A performant ahead-of-time compiled e-graph engine
01h 03mScaling up machine learning without tears (and what do programming languages ha...
15mTool Talk: Soufflé
30mA Compiler-First Planetary Compute Engine: Automatic differentiable and performance(…)
14mAchieving High-Performance the Functional Way - A Functional Pearl on Expressing High...
30mGenerating High Performance Code for Irregular Data Structures using Dependent Types
15mFormulog: Datalog for SMT-Based Static Analysis
13mMind the Abstraction Gap: Bringing Equality Saturation to Real-World ML Compilers