ICL Research Profile

Overview

The Sparse direct methods via Run-time Scheduling and Execution of Kernels with Auto-tunable and Frequency-scaling Features for Energy-aware computing on heterogeneous architectures (SparseKaffe) project creates fast and efficient sparse direct methods for platforms with multi-core processors with one or more accelerators (e.g., GPUs or Xeon Phi coprocessors). SparseKaffe spans the platform pyramid, from desktop machines to extreme-scale systems consisting of multiple heterogeneous nodes connected through a high-speed network, with the goal of achieving orders of magnitude gains in computational performance while also paying careful attention to energy requirements.

The SparseKaffe project is a collaboration between UTK, the University of Florida, and Texas A&M University. ICL’s work on the project concentrates on kernel designs and performance tuning, as well as on dynamic runtime scheduling using a dataflow model. This work will leverage—and be a natural extension of—ICL’s work on runtimes as part of the MAGMA, PLASMA, and PaRSEC projects. The autotuning of the algorithm-specific computational kernels will apply the principles behind ICL’s DARE project.