Partitioning Models for General Medium-Grain Parallel Sparse Tensor Decomposition

The focus of this article is efficient parallelization the canonical polyadic decomposition algorithm utilizing alternating least squares method for sparse tensors on distributed-memory architectures. We propose a hypergraph model general medium-grain partitioning which does not enforce any topological constraint partitioning. proposed based splitting given tensor into nonzero-disjoint component tensors. Then mode-dependent coarse-grain constructed each tensor. A net amalgamation operation to form composite from these hypergraphs correctly encapsulate minimization communication volume. heuristic splits nonzeros dense slices obtain in So we partially attain slice coherency at (sub)slice level since performed (sub)slices instead individual nonzeros. also utilize well-known recursive-bipartitioning framework improve quality heuristic. Finally, tripartite graph with aim faster expense increasing total Parallel experiments conducted 10 real-world up 1024 processors confirm validity and models.