Impact of Data Distribution and Schedulers for the LU Factorization on Multi-Core Clusters

• Otho Jose Sirtoli Marcondes UFRGS
• Philippe Olivier Alexandre Navaux UFRGS
• Lucas Mello Schnorr UFRGS

Resumo

The performance of dense linear algebra operations on multi-core clusters depends critically on the interaction between data distribution strategies and task scheduling policies. This paper investigates the trade-offs between different block-cyclic (BC) data partitioning configurations and intra-node scheduling heuristics for tiled LU factorization on small-scale HPC clusters. We conduct experiments on a 144-core cluster (six nodes with 2×12 cores each) using the CHAMELEON dense linear algebra library with StarPU-MPI runtime system. Our study examines how different scheduling heuristics influence overall execution time. Through systematic evaluation of data partitioning configurations (1×6, 2×3, 3×2, and 6×1) and comprehensive trace analysis using the StarVZ visualization framework, we provide insights into optimal tile sizing, correlations between data partitioning and scheduling strategies, and methodological considerations for reproducible HPC experimentation. Our findings demonstrate the complex interplay between static data distribution choices and dynamic scheduling decisions, offering practical guidance for optimizing dense linear algebra computations on modern multi-core cluster architectures.

Palavras-chave: Systematics, Runtime, Processor scheduling, Multicore processing, Distribution strategy, High performance computing, Linear algebra, Computer architecture, Dynamic scheduling, Libraries, homogeneous clusters, lu factorization, data partitioning, task-based applications