[Paper Review] Partitioning Strategies for Load Balancing Subgraph-centric Distributed Graph Processing

Neel Choudhury, Ravikant Dindokar+2 more|arXiv (Cornell University)|Aug 18, 2015

Graph Theory and Algorithms24 references3 citations

TL;DR

This paper proposes three subgraph-centric partitioning strategies—flat, hierarchical, and adaptive—to improve load balancing in distributed graph processing. By optimizing component distribution across machines, the approach boosts CPU utilization by up to 8× and reduces makespan by up to 5× compared to default strategies, outperforming Apache Giraph in scalability and efficiency on power-law and spatio-planar graphs.

ABSTRACT

Distributed graph processing platforms have helped emerging application domains use commodity clusters and Clouds to analyze large graphs. Vertex-centric programming models like Google Pregel, and their subgraph-centric variants, specify data-parallel application logic for a single vertex or component that execute iteratively. The locality and balancing of components within partitions affects the performance of such platforms. We propose three partitioning strategies for a subgraph-centric model, and analyze their impact on CPU utilization, communication, iterations, and makespan. We analyze these using Breadth First Search and PageRank algorithms on powerlaw and spatio-planar graphs. They are validated on a commodity cluster using our GoFFish subgraph-centric platform, and compared against Apache Giraph vertex-centric platform. Our experiments show upto 8 times improvement in utilization resulting to upto 5 times improvement of overall makespan for flat and hierarchical partitioning over the default strategy due to improved machine utilization. Further, these also exhibit better horizontal scalability relative to Giraph.

Motivation & Objective

• To address performance bottlenecks in distributed graph processing caused by poor load balancing and communication overhead.
• To improve CPU utilization and reduce makespan in subgraph-centric graph processing platforms.
• To evaluate partitioning strategies that enhance horizontal scalability and resource utilization on commodity clusters.
• To compare subgraph-centric partitioning against vertex-centric models like Apache Giraph in real-world workloads.

Proposed method

• Proposes flat partitioning, which divides the graph into subgraphs based on vertex degree and connectivity to balance work per machine.
• Introduces hierarchical partitioning, which recursively subdivides large subgraphs to reduce load skew and improve load distribution.
• Designs adaptive partitioning that dynamically reassigns subgraphs based on runtime performance metrics to maintain balance.
• Employs the GoFFish platform to implement and evaluate subgraph-centric execution with iterative algorithms like BFS and PageRank.
• Uses power-law and spatio-planar graphs to simulate real-world workloads and measure performance across CPU utilization, communication, iterations, and makespan.
• Compares results against Apache Giraph’s vertex-centric model to assess scalability and efficiency improvements.

Experimental results

Research questions

• RQ1 How do different subgraph partitioning strategies affect CPU utilization in distributed graph processing?
• RQ2 To what extent can subgraph-centric partitioning reduce makespan compared to default and vertex-centric strategies?
• RQ3 How does the proposed partitioning improve horizontal scalability on commodity clusters?
• RQ4 What is the impact of partitioning on communication volume and iteration count in iterative graph algorithms?

Key findings

• Flat and hierarchical partitioning improved CPU utilization by up to 8 times compared to the default strategy.
• Makespan was reduced by up to 5 times using the proposed partitioning strategies over the default approach.
• The subgraph-centric approach demonstrated better horizontal scalability than Apache Giraph across all tested graph types.
• Breadth-First Search and PageRank workloads showed significant performance gains due to reduced load imbalance and communication overhead.
• Adaptive partitioning maintained high load balance during execution, leading to consistent performance improvements.