Minisymposium Presentation

Algebraic Programming, or ALP for short, allows for programming with explicit algebraic structures. Such structures range from the simplistic such as associative binary operators, to richer constructs such as semirings. This algebraic knowledge, given to ALP by the programmer, percolates through the ALP framework and allows it to auto-parallelise, as well as perform other types of automatic program transformations for achieving high performance.Originating from the GraphBLAS, ALP’s initial and most mature interface concerns generalized linear algebra. This talk focuses on two recent works related to the use of ALP/GraphBLAS within high-performance machine learning. First, we shall overview how the ALP framework achieves interoperability with Spark, making accessible ALP/GraphBLAS algorithms from within Spark, a standard framework for Big Data analytics. Second, the talk details the ALP/Pregel implementation, which relies on ALP/GraphBLAS to simulate, at high efficiency and at full scalability, vertex-centric programming.Performance comparisons of canonical machine learning algorithms such as PageRank versus the standard GraphX library on top of Spark display significant performance gains of up to 21x on two nodes. The vertex-centric ALP/Pregel furthermore is shown to auto-parallelise well on shared-memory architectures; a vertex-centric PageRank-like algorithm achieves a 5.7x speedup versus a highly optimised parallel linear algebraic PageRank.