HARP-ME GPU framework speeds up graph motif enumeration 10x
New GPU technique reduces candidate expansion with closure-aware compilation and reuse...
Deep Dive
HARP-ME, a GPU framework for exact induced 4-node motif enumeration, introduces closure-aware compilation and induced-signature reuse to reduce repeated set intersections. Across six graphs, it is up to 2.11× faster than Pangolin, 1.83× over partitioned PBE, and 10.73× over the evaluated CPU baseline.
Key Points
- Introduces closure-aware compilation to select optimal anchors based on cost, derivation benefit, state reuse, and halo overhead.
- Achieves up to 2.11× speedup over Pangolin and 10.73× over CPU baselines on 6 graph types.
- Uses induced-signature reuse and canonical anchor-owner rule for exact counting across GPU memory partitions.
Why It Matters
Enables faster graph mining on GPUs, crucial for analyzing complex networks in social media, biology, and security.