Home

This software stack provides sketching-based NLA kernels for more general data analysis and optimization applications; such tasks have significantly different input matrices and performance criteria than arise in the more traditional scientific computing applications. The crucial NLA kernels to be implemented include regression and low-rank approximations of matrices, akin to the Singular Value Decomposition (SVD).

Additionally this library provides a simple distributed Python interface.

---

Elemental

Skylark uses Elemental for a dense matrix functionality. Elemental is a framework for distributed-memory dense linear algebra that strives to be both fast and convenient. It combines ideas including: element-wise matrix distributions (Hedrickson et al.), object-oriented submatrix tracking (FLAME, van de Geijn et al.), and first-class matrix distributions (PLAPACK, van de Geijn et al.). Many algorithms use techniques from LAPACK (Anderson et al.) in order to improve numerical stability.

• Elemental
• A tutorial on the design and usage of Elemental
• TOMS paper
• Elemental Distributions & Matrix Multiplication Overview paper

CombBLAS

Skylark uses CombBLAS for sparse matrix functionality. CombBLAS is a distributed memory reference implementation that implements scalable sparse (and some dense) matrix operations that is used to implement graph algorithms such as betweenness centrality and Markov clustering, see:

• CombBLAS (Doxygen)
• Presentation
• KDT (contains SWIG bindings for CombBLAS)
• KDT Installation Guide