PulseExploreJournal ClubDebatesTrendingResearchersJournals
Instagram
HomeExploreJournal ClubTrending
Synapse
⌘+K
Synapse
April 25, 2026ACM SIGMOD Record0 citationsOpen Access

One Graph, Many Sources: Automating Vectorization Safely

WZWenyue ZhaoYCYang CaoPBPeter Buneman

Key Points

  • To develop a framework that safely automates vectorization for graph workloads to achieve higher performance while ensuring correctness.
  • Introduced AutoMI, a source-to-source compilation framework for vectorization.
  • Used a per-message bit-vector to track per-source activation for SIMD updates.
  • Implemented TrackFree optimization to streamline code when over-activation is not problematic.
  • Achieved 9.6 to 29.5x speedup over parallelized serial evaluation.
  • Obtained 7.1 to 26.4x speedup compared to batch processing.
  • Outperformed prior multi-instance code by 2.6 to 4.6x.

Abstract

Many graph workloads repeatedly run the same traversal or iterative computation from many different source vertices. Multi-instance execution can share work across sources and, if implemented carefully, exploit SIMD by packing per-source state into vectors. However, in vertex-centric graph processing systems, naive ''vectorize everything'' transformations can be wrong: a vertex may be active for only a subset of sources in a round, yet unguarded SIMD updates implicitly advance all lanes, leading to incorrect answers. We present AutoMI, a source-to-source compilation framework that automatically converts a single-instance program into a provably correct SIMD-vectorized multi-instance program runnable on existing vertex-centric engines. AutoMI uses a per-message bit-vector track to recover per-source activation and generate masked SIMD updates. It further provides a TrackFree optimization, guided by an algebraic idempotence characterization, that safely drops track and produces streamlined code when over-activation is harmless. On six real graphs (up to billions of edges) and up to 256 sources, AutoMI achieves 9.6 to 29.5x speedup over parallelized serial evaluation, 7.1 to 26.4x over batch processing, and 2.6 to 4.6x over prior handcrafted multi-instance code.

Ask AI
Helpful
Bookmark
Share
View Full Paper

Cite This Study

Zhao et al. (2026) studied this question.

synapsesocial.com/papers/69ec5ac988ba6daa22dac513https://doi.org/10.1145/3810900.3810904
Ask AI
Helpful
Bookmark
Share
View Full Paper