RingSampler: GNN Sampling on Large-Scale Graphs with io_uring (HotStorage 2025)
·82 words·1 min
Table of Contents
Overview #
RingSampler leverages Linux io_uring for highly parallel out-of-core neighborhood sampling on graphs that exceed memory capacity. The design saturates storage and CPU resources to narrow the gap to in-memory sampling.
Key Ideas #
- Asynchronous, batched I/O via io_uring to minimize syscalls and context switches.
- Careful batching and memory layout to improve cache behavior and reduce copies.
- End-to-end pipeline overlapping compute and I/O per minibatch.
Results #
Achieves near in-memory sampling performance while operating fully out-of-core in our evaluations.