G-MAP: A Graph-Neural Network Based Framework for Memory Access Prediction
Published in IEEE HPEC, 2023
Recommended citation: A. R. Gorle, P. Zhang, R. Kannan and V. K. Prasanna, "G-MAP: A Graph Neural Network-Based Framework for Memory Access Prediction," 2023 IEEE High Performance Extreme Computing Conference (HPEC), Boston, MA, USA, 2023, pp. 1-7, doi: 10.1109/HPEC58863.2023.10363605. https://ieeexplore.ieee.org/document/10363605
Memory access prediction is a crucial problem in data prefetchers, as it helps us improve memory performance and reduce latency in computing systems. Existing works model the problem as a sequence prediction problem. This can be limited in its ability to capture complex patterns and dependencies in memory access behavior. In recent years, Graph Neural Networks (GNNs) have emerged as a promising technique for modeling and predicting complex relationships in graph-structured data.
In this paper, we introduce G-MAP, a novel Graph Neural Network-based framework for Memory Access Prediction. First, we propose Mem2Graph, a novel approach mapping a memory access sequence to a graph representation, capturing both the spatial and temporal locality in the sequence. Second, we implement various GNNs for G-MAP, including Graph Convolutional Network (GCN), Gated Graph Sequence Neural Network (GG-NN), and Graph Attention Network (GAT). Those models take the graph generated from Mem2Graph as input and predict future memory address jumps (deltas). We evaluate the effectiveness of G-MAP using the SPEC 2006 benchmark. G-MAP using GG-NN shows the highest among all models, achieving averagely 0.7526 F1-Score, which is 10.77% higher than the Multi-Layer Perceptron baseline.
Future work:
- Building a weighted graph representation, making the mapping more robust.
- Implemented state-of-the-art second models for the graph representation like GINs, GraphSage, and GraphSaint.
- Extending this technique to a secondary storage system instead of just cache memory.
- Heterogenous GNN for the same task, and adding time step to the nodes for a more dynamic representation.