Pluto: High-Performance, Memory-Efficient Distributed Graph Analytics through Advanced Mirroring

Inter-host communication poses a significant performance bottleneck in distributed graph analytics due to synchronization overheads. To mitigate this, state-of-the-art systems typically employ the full mirroring technique to replicate all potentially needed remote data and adopt a bulk-synchronous parallel execution model for coarse-grained communication. While effective for reducing network traffic, these approaches substantially increase memory footprint and constrain system parallelism. This paper introduces Pluto, a memory-efficient distributed graph analytics system based on two advanced mirroring designs: static partial mirroring and a mirror-free architecture. Nonproductive data duplication is avoided to reduce memory usage, while the work migration mechanism allows communication-computation overlap for performance improvement. For homogeneous graphs, Pluto achieves up to 3.8× speedup (harmonic mean 1.75×) compared to a full mirroring baseline and delivers up to 12× speedup (harmonic mean 1.75×) over existing open-source systems. For labeled property graphs, Pluto achieves up to 2.6× speedup (harmonic mean 1.37×) and lowers the minimum host requirement to 50%–90% of the baseline.