Workshop: 17th Workshop on Latest Advances in Scalable Algorithms for Large-Scale Heterogeneous Systems (ScalAH'26)
Authors: Mohammad Alaul Haque Monil (Oak Ridge National Laboratory (ORNL)), Sanil Rao (Carnegie Mellon University), Het Mankad and Keita Teranishi (Oak Ridge National Laboratory), Franz Franchetti (Carnegie Mellon University), and Jeffrey Vetter (Oak Ridge National Laboratory)
Abstract: Contemporary HPC, cloud, and embedded systems combine heterogeneous processors, and using them effectively requires careful trade-offs among kernel implementation, multi-device orchestration, and data transfers. Domain scientists need a framework that dynamically discovers performance-portable solutions from the characteristics of the application and the underlying multi-accelerator platform. We propose IRISX, which integrates the SPIRAL code generation system with the IRIS runtime to deploy architecture-optimized kernels together with a portable abstraction for multi-device heterogeneity. IRISX transforms computation at the kernel, task, and graph levels—through kernel, task, and DAG fusion—to fit each target ar- chitecture, and a hybrid analytical and ML model selects near-optimal configurations in 96% of unseen cases. On a partial differential equation library evaluated across Frontier, Aurora, and multiple ExCL systems, IRISX delivers up to 25.9× speedup over unoptimized execution, up to 10× over the Proto library, and up to 74× over a StarPU implementation.
Back to 17th Workshop on Latest Advances in Scalable Algorithms for Large-Scale Heterogeneous Systems (ScalAH'26) Archive Listing
Back to Full Workshop Archive Listing