Scalable parallel building blocks for custom data analysis

Abstract

We present a set of building blocks that provide scalable data movement capability to computational scientists and visualization researchers for writing their own parallel analysis. The set includes scalable tools for domain decomposition, process assignment, parallel I/O, global reduction, and local neighborhood communication-tasks that are common across many analysis applications. The global reduction is performed with a new algorithm, described in this paper, that efficiently merges blocks of analysis results into a smaller number of larger blocks. The merging is configurable in the number of blocks that are reduced in each round, the number of rounds, and the total number of resulting blocks. We highlight the use of our library in two analysis applications: parallel streamline generation and parallel Morse-Smale topological analysis. The first case uses an existing local neighborhood communication algorithm, whereas the latter uses the new merge algorithm.

Block-parallel data analysis with DIY2

“…It is the job of the runtime to map those instructions to code running on a process and messages exchanged between processes. The starting point of our work is DIY1 [32,34], a C library that structures data into blocks. It expects the computation to be organized in a bulk-synchronous pattern, but does not enforce this structure through programming convention.…”

“…We compare the performance with DIY1 [32,34] (which only supports in-core blocks) and when possible, with equivalent collective functions in MPI itself (by assigning one block per MPI rank). We want to Figure 2: Cian mini-app merge-and swap-reduce using DIY2 compared with MPI using reduce and reducescatter, respectively.…”

“…Topological structures such as Morse-Smale complexes [16] and persistence diagrams [23] can be reduced this way. The merge-reduction algorithm in DIY1 was first published in 2011 [34]. The swap communication pattern is used for associative reduction of homogeneous contiguous data buffers that remain distributed instead of being merged into a smaller number of blocks.…”

Parallel ptychographic reconstruction

“…It also uses the DIY [23] data-parallel programming library for the multi-GPU extension, the HDF5 library for data I/O, and the Simple DirectMedia Layer (SDL) [24] library for OpenGL visualization. The programming language used within CUDA (CUDA-C) is an extension of the C programming language which allows one to implement GPU-based parallel functions, called kernels, which, when called, are executed n times in parallel by n different CUDA threads.…”

“…Given any of these patterns, spatially contiguous scan subregions can be defined such that the degree of overlap between adjacent scan points is preserved. Data partitioning is achieved using the DIY parallel programming library [23] that is written on top of MPI to facilitate communication between parallel processes. In DIY terminology, we assign a DIY block (not to be confused with CUDA blocks) to each GPU.…”