关键词:分布式库张量;线性代数矩阵乘法;载平衡;细分框架
摘 要:Cyclops (cyclic-operations) Tensor Framework (CTF) 1 isa distributed library for tensor contractions. CTF aims toscale high-dimensional tensor contractions done in CoupledCluster calculations on massively-parallel supercomputers.The framework preserves tensor symmetry by subdividingtensors cyclically, producing a highly regular parallel decomposition.The parallel decomposition e ectively hidesany high dimensional structure of tensors reducing the complexityof the distributed contraction algorithm to knownlinear algebra methods for matrix multiplication. We alsodetail the automatic topology-aware mapping framework deployedby CTF, which maps tensors of any dimension andstructure onto torus networks of any dimension. We employ;virtualization to provide completely general mapping support while maintaining perfect load balance. Performance of a preliminary version of CTF on the IBM Blue Gene/P and Cray XE6 supercomputers shows highly ecient weakscaling, demonstrating the viability of our approach.