GraphMat: High performance graph analytics made productive (1503.07241v1)

Abstract: Given the growing importance of large-scale graph analytics, there is a need to improve the performance of graph analysis frameworks without compromising on productivity. GraphMat is our solution to bridge this gap between a user-friendly graph analytics framework and native, hand-optimized code. GraphMat functions by taking vertex programs and mapping them to high performance sparse matrix operations in the backend. We get the productivity benefits of a vertex programming framework without sacrificing performance. GraphMat is in C++, and we have been able to write a diverse set of graph algorithms in this framework with the same effort compared to other vertex programming frameworks. GraphMat performs 1.2-7X faster than high performance frameworks such as GraphLab, CombBLAS and Galois. It achieves better multicore scalability (13-15X on 24 cores) than other frameworks and is 1.2X off native, hand-optimized code on a variety of different graph algorithms. Since GraphMat performance depends mainly on a few scalable and well-understood sparse matrix operations, GraphMatcan naturally benefit from the trend of increasing parallelism on future hardware.

• The paper demonstrates that GraphMat bridges the productivity-performance gap by mapping vertex programs to sparse matrix operations, achieving near-native performance.
• The paper reports a 13-15X speedup on multicore systems and outperforms GraphLab by up to 7.9X in diverse graph analytics tasks.
• The paper shows that refining backend efficiencies via sparse matrix-vector multiplications simplifies high-performance graph analytics for developers.

An Overview of GraphMat: Advancements in High-Performance Graph Analytics

The paper "GraphMat: High performance graph analytics made productive" presents an innovative approach to bridging the gap between productivity and performance in large-scale graph analytics. The authors introduce GraphMat, a framework designed to map vertex programs to high-performance sparse matrix operations, leveraging the advantages of both vertex programming approaches and matrix-oriented backend efficiencies.

Main Contributions

GraphMat presents several notable contributions to the field of graph analytics:

1. Near-Native Performance: GraphMat achieves performance levels within 1.2X of hand-optimized native code across diverse graph algorithms. This is significant given that the typical challenge in graph frameworks is the order-of-magnitude performance gap between user-friendly and highly optimized code.
2. Scalability: The framework demonstrates enhanced multicore scalability, achieving a 13-15X speedup on 24-core systems. This scalability surpasses that of existing frameworks, namely GraphLab, CombBLAS, and Galois, which exhibit lower scalability.
3. Productivity: Users benefit from the simplicity of a vertex programming model without facing the steep performance penalties usually associated with such models. GraphMat abstracts the complexity of sparse matrix operations, making it accessible to developers familiar with vertex-centric programming.
4. Backend Efficiencies: By employing a sparse matrix-vector multiplication paradigm, GraphMat optimizes backend operations, allowing it to take advantage of established research in high-performance computing.

Numerical Results and Performance

GraphMat exhibits significant runtime improvements for various graph algorithms and real-world datasets. The framework outperforms GraphLab by a factor of 7.5X for PageRank and 7.9X for BFS, indicating its robust backend optimizations. Similarly, it surpasses CombBLAS by an average of 2-4X, highlighting its superior code efficiency and architectural alignment. Compared to Galois, GraphMat exhibits a slight performance edge, with an average speedup of 1.2X across applications, despite Galois’ strong task-based design.

The paper specifies that GraphMat's performance benefits stem from reduced instruction counts, decreased stall cycles, and optimal utilization of memory bandwidth. These aspects collectively embody the core strengths of the framework's design.

Theoretical and Practical Implications

Theoretically, GraphMat reinforces the feasibility of using sparse linear algebra as a conduit to optimize graph analytics. The paper bridges the knowledge and techniques from high-performance computing with the demands of large-scale data processing in graphs, encouraging future frameworks to consider similar architectural paths.

Practically, GraphMat provides a potent template for developing scalable graph analytics frameworks that do not sacrifice productivity for performance. In operational settings, this means organizations can achieve closer-to-native code performance without necessitating specialized development efforts aligned with hardware intricacies.

Future Directions

The authors suggest that the future trajectory for GraphMat includes extending its efficient single-node performance to distributed environments. Given the framework's basis in well-understood sparse matrix operations, it is poised to naturally extend to multi-node and heterogeneous computing settings, including GPU and co-processor architectures.

Furthermore, the integration of GraphMat's optimizations into existing frameworks like CombBLAS presents opportunities for broader application across different programming models. Overall, GraphMat paves the way for more productive and performant graph analytics, aligning closely with advancements in array-based and high-performance computing systems.

In conclusion, GraphMat provides an exemplary model for enhancing graph analytics by marrying productivity with high-performance computing techniques, setting a noteworthy precedent for future research and development in the discipline.