理化学研究所 計算科学研究センター

メニュー
メニュー

第154回

第154回
日時: 2018年11月27日(火)、14:00 - 15:00
場所: R-CCS 6階講堂

・講演題目:Performance portable parallel CP-APR tensor decompositions
・講演者:寺西 慶太(Principal Member of Technical Staff, Sandia National Laboratories, California)
※発表・スライド共に英語

講演要旨: 詳細を見る

Tensors have found utility in a wide range of applications, such as chemometrics, network traffic analysis, neuroscience, and signal processing. Many of these data science applications have increasingly large amounts of data to process and require high-performance methods to provide a reasonable turnaround time for analysts. Sparse tensor decomposition is a tool that allows analysts to explore a compact representation (low-rank models) of high-dimensional data sets, expose patterns that may not be apparent in the raw data, and extract useful information from the large amount of initial data. In this work, we consider decomposition of sparse count data using CANDECOMP-PARAFAC Alternating Poisson Regression (CP-APR).
Unlike the Alternating Least Square (ALS) version, CP-APR algorithm involves non-trivial constraint optimization of nonlinear and nonconvex function, which contributes to the slow adaptation to high performance computing (HPC) systems. The recent studies by Kolda et al. suggest multiple variants of CP-APR algorithms amenable to data and task parallelism together, but their parallel implementation involves several challenges due to the continuing trend toward a wide variety HPC system architecture and its programming models.
To this end, we have implemented a production-quality sparse tensor decomposition code, named SparTen, in C++ using Kokkos as a hardware abstraction layer. By using Kokkos, we have been able to develop a single code base and achieve good performance on each architecture. Additionally, SparTen is templated on several data types that allow for the use of mixed precision to allow the user to tune performance and accuracy for specific applications. In this presentation, we will use SparTen as a case study to document the performance gains, performance/accuracy tradeoffs of mixed precision in this application, development effort, and discuss the level of performance portability achieved. Performance profiling results from each of these architectures will be shared to highlight difficulties of efficiently processing sparse, unstructured data. By combining these results with an analysis of each hardware architecture, we will discuss some insights for improved use of the available cache hierarchy, potential costs/benefits of analyzing the underlying sparsity pattern of the input data as a preprocessing step, critical aspects of these hardware architectures that allow for improved performance in sparse tensor applications, and where remaining performance may still have been left on the table due to having single algorithm implementations on diverging hardware architectures.