Guided installation of basic linear algebra routines in a clusterrnwith manycore components

摘要

Computational systems are nowadays composed of basic computational components that sharernmultiprocessors and coprocessors of different types, typically several graphics processing unitsrn(GPUs) or many integrated cores (MICs), and those computational components are combined inrnheterogeneous clusters of nodes with different characteristics, including coprocessors of differentrntypes, with varying numbers of nodes at different speeds. The software previously developedrnand optimized for simpler system needs to be redesigned and reoptimized for these new, morerncomplex systems. The adaptation to hybrid multicore + multiGPU and multicore + multiMIC ofrnautotuning techniques for basic linear algebra routines is analyzed. The matrix-matrixmultiplicationrnkernel, which is optimized for different computational system components through guidedrnexperimentation, is studied. The routine is installed for each node in the cluster, and the informationrngenerated from individual installations may be used for a hierarchical installation in a cluster.rnThe basic matrix-matrix multiplication may, in turn, be used inside higher level routines, whichrndelegate their efficient execution to the optimization of the lower level routine. Experimentalrnresults are satisfactory in different multicore + multiGPU and multicore + multiMIC systems. Sornthe guided search of execution configurations for satisfactory execution times proves to be a usefulrntool for heterogeneous systems, where the complexity of the system means a correct use ofrnhighly efficient routines and libraries is difficult.