Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

GPU-Accelerated Algorithms for Process Mapping

Petr Samoldekin, Christian Schulz, Henning Woydt

TL;DR

The paper tackles the GPMP by introducing two GPU-accelerated strategies: a hierarchical multisection method that traverses the system hierarchy with GPU-based partitioning, and an integrated mapping approach that weaves the mapping objective into a multilevel partitioning pipeline. Both methods exploit GPU parallelism (via Kokkos and CSR-based data layouts) to substantially reduce runtimes while targeting effective communication costs, with GPU-HM delivering strong quality and GPU-IM achieving exceptional speedups at some quality trade-offs. Experimental results show speedups well over CPU baselines (including state-of-the-art SharedMap and IntegratedMapping) and demonstrate the practical potential of GPU-driven process mapping for large-scale HPC topologies. The work demonstrates the feasibility of first GPU-based process-mapping solutions and sets the stage for further improvements in refinement strategies, coarsening, and emerging GPU partitioners. The significance lies in enabling rapid, scalable mapping decisions that directly impact application performance on modern hierarchical supercomputers.

Abstract

Process mapping asks to assign vertices of a task graph to processing elements of a supercomputer such that the computational workload is balanced while the communication cost is minimized. Motivated by the recent success of GPU-based graph partitioners, we propose two GPU-accelerated algorithms for this optimization problem. The first algorithm employs hierarchical multisection, which partitions the task graph alongside the hierarchy of the supercomputer. The method utilizes GPU-based graph partitioners to accelerate the mapping process. The second algorithm integrates process mapping directly into the modern multilevel graph partitioning pipeline. Vital phases like coarsening and refinement are accelerated by exploiting the parallelism of GPUs. In our experiments, both methods achieve speedups exceeding 300 when compared to state-of-the-art CPU-based algorithms. The first algorithm has, on average, about 10 percent greater communication costs and thus remains competitive to CPU algorithms. The second approach is much faster, with a geometric mean speedup of 77.6 and peak speedup of 598 at the cost of lower solution quality. To our knowledge, these are the first GPU-based algorithms for process mapping.

GPU-Accelerated Algorithms for Process Mapping

TL;DR

The paper tackles the GPMP by introducing two GPU-accelerated strategies: a hierarchical multisection method that traverses the system hierarchy with GPU-based partitioning, and an integrated mapping approach that weaves the mapping objective into a multilevel partitioning pipeline. Both methods exploit GPU parallelism (via Kokkos and CSR-based data layouts) to substantially reduce runtimes while targeting effective communication costs, with GPU-HM delivering strong quality and GPU-IM achieving exceptional speedups at some quality trade-offs. Experimental results show speedups well over CPU baselines (including state-of-the-art SharedMap and IntegratedMapping) and demonstrate the practical potential of GPU-driven process mapping for large-scale HPC topologies. The work demonstrates the feasibility of first GPU-based process-mapping solutions and sets the stage for further improvements in refinement strategies, coarsening, and emerging GPU partitioners. The significance lies in enabling rapid, scalable mapping decisions that directly impact application performance on modern hierarchical supercomputers.

Abstract

Process mapping asks to assign vertices of a task graph to processing elements of a supercomputer such that the computational workload is balanced while the communication cost is minimized. Motivated by the recent success of GPU-based graph partitioners, we propose two GPU-accelerated algorithms for this optimization problem. The first algorithm employs hierarchical multisection, which partitions the task graph alongside the hierarchy of the supercomputer. The method utilizes GPU-based graph partitioners to accelerate the mapping process. The second algorithm integrates process mapping directly into the modern multilevel graph partitioning pipeline. Vital phases like coarsening and refinement are accelerated by exploiting the parallelism of GPUs. In our experiments, both methods achieve speedups exceeding 300 when compared to state-of-the-art CPU-based algorithms. The first algorithm has, on average, about 10 percent greater communication costs and thus remains competitive to CPU algorithms. The second approach is much faster, with a geometric mean speedup of 77.6 and peak speedup of 598 at the cost of lower solution quality. To our knowledge, these are the first GPU-based algorithms for process mapping.
Paper Structure (15 sections, 3 equations, 2 figures, 1 table, 4 algorithms)

This paper contains 15 sections, 3 equations, 2 figures, 1 table, 4 algorithms.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. Concepts and Notations
  4. Multilevel Graph Partitioning
  5. Related Work
  6. GPU Multilevel Graph Partitioning
  7. Process Mapping
  8. Kokkos
  9. CSR Graph Format
  10. GPU-Accelerated Process Mapping
  11. Hierarchical Multisection
  12. Integrated Mapping
  13. Experimental Evaluation
  14. Results
  15. Conclusion

Figures (2)

  • Figure 1: Comparison of solution quality (a) and speedup (b) of our proposed algorithms. Table (c) shows the runtime breakdown per phase.
  • Figure 2: Comparison of speedup (left) and performance profile of achieved communication costs (right) of our proposed algorithms GPU-HM-ultra and GPU-IM to CPU-based algorithms SharedMap and IntegratedMapping.