FlashX is a collection of big data analytics tools that perform data analytics in the form of graphs and matrices. They utilize solid-state drives (SSDs) to scale to large datasets in a single machine. It has five main components: SAFS, FlashGraph, FlashMatrix, FlashEigen and FlashR.
SAFS a user-space filesystem designed for large SSD arrays. It is capable of achieving the maximal I/O throughput from SSD arrays (millions of I/O per second for random I/O and tens of gigabytes per second for sequential I/O). The I/O performance is further magnified by the light-weight page cache if the workload can generate cache hits.
FlashGraph is a general-purpose graph analysis framework that exposes vertex-centric programming interface for users to express varieties of graph algorithms. FlashGraph scales graph computation to large graphs by keeping the edges of a graph on SSDs and computation state in memory. With smart I/O scheduling, FlashGraph is able to achieve performance comparable to state-of-art in-memory graph analysis frameworks and significantly outperforms state-of-art distributed graph analysis frameworks while being able to scale to graphs with billions of vertices and hundreds of billions of edges. Please see the performance result.
FlashMatrix is a matrix computation engine that provides a small set of generalized matrix operations on sparse matrices and dense matrices to express varieties of data mining and machine learning algorithms. For certain graph algorithms such as PageRank, which can be formulated as sparse matrix multiplication, FlashMatrix is able to significantly outperform FlashGraph.
FlashEigen is an eigensolver that extends the Anasazi eigensolvers with FlashMatrix. It computes eigenvalues of billion-node graphs efficiently and pricisely in a single machine. As such, FlashEigen enables users to perform spectral analysis on very large graphs in a single machine.
FlashR extends the existing R programming framework to process datasets at a scale of terabytes. FlashR integrates the matrices and generalized operators of FlashMatrix with R so that R users can implement many data mining and machine learning algorithms completely in R with performance comparable to optimized C implementations. FlashR reimplements many existing R functions for matrices to provide users a familiar R programming environment. FlashR is implemented as a regular R package. The programming tutorial shows all of the features in FlashR.
FlashGraph programming tutorial.
FlashX performance and scalability
Da Zheng, Disa Mhembere, Joshua T. Vogelstein, Carey E. Priebe, and Randal Burns, “FlashMatrix: Parallel, scalable data analysis with generalized matrix operations using commodity ssds,” arXiv preprint arXiv:1604.06414, 2016 [pdf]
Da Zheng, Disa Mhembere, Vince Lyzinski, Joshua Vogelstein, Carey E. Priebe, and Randal Burns, “Semi-external memory sparse matrix multiplication on billion-node graphs”, Transactions on Parallel and Distributed Systems, 2016. [pdf]
Heng Wang, Da Zheng, Randal Burns, Carey Priebe, Active Community Detection in Massive Graphs, SDM-Networks 2015 [pdf]
Da Zheng, Disa Mhembere, Randal Burns, Joshua Vogelstein, Carey E. Priebe, Alexander S. Szalay, FlashGraph: Processing Billion-Node Graphs on an Array of Commodity SSDs, FAST'15, [pdf][bib]
Da Zheng, Randal Burns, Alexander S. Szalay, Toward Millions of File System IOPS on Low-Cost, Commodity Hardware, in Proceeding SC '13 Proceedings of the International Conference on High Performance Computing, Networking, Storage and Analysis, [pdf][bib]
Mailing list: [email protected]
