Code Monkey home page Code Monkey logo

bspmm's Introduction

NWChem's BSPMM communication pattern

Block sparse matrix multiplication (BSPMM) is the dominant cost in the CCSD and CCSD(T) quantum chemical many-body methods of NWChem, a prominent quantum chemistry application suite for large-scale simulations of chemical and biological systems. NWChem implements its BSPMM with dense matrix operations using a get-compute-update pattern: each worker (processing entity) uses MPI_Get to retrieve the submatrices it needs, and after the multiplication it uses an MPI_Accumulate to update the memory at the target location.

This repository contains mini-apps that implement a 2D version of BSPMM to perform A ร— B = C, wherein the input matrices A and B are composed of tiles. The nonzero tiles are evenly distributed among the ranks in a round-robin fashion. Each rank maintains a work-unit table that lists all the multiplication operations that workers need in order to cooperatively execute. Rank 0 hosts a global counter, which the workers fetch and add atomically (MPI_Fetch_and_op). The fetched counter serves as an index to the work-unit table. Each worker locally accumulates its C tiles until the next fetched work unit corresponds to a different C tile, in which case the worker uses an MPI_Accumulate to update the C tile. A worker is a process in MPI everywhere and a thread in MPI+threads.

This mini-app was written after discussions with Pavan Balaji, Min Si, and Shintaro Iwasaki from Argonne National Laboratory, and with Jeff Hammond from Intel.

Versions

bspmm_single.c

  • MPI everywhere version (flat MPI)

bspmm_multiple.c

  • MPI+OpenMP version using MPI_THREAD_MULTIPLE
  • Expressing no logical parallelism exposed to the MPI library

bspmm_multiple_nwins.c

  • MPI+OpenMP version using MPI_THREAD_MULTIPLE
  • Expressing logical parallelism to the MPI library
    • The MPI_Get operations of each thread are independent. Hence, each thread uses their own window.
    • All threads must use the same window for MPI_Accumulate operations since atomicity across windows for the same memory location is undefined. However, the issue of MPI_Accumulate operations from multiple threads could still occur independently since BSPMM does not require ordering of these operations. Hence, we hint accumulate_ordering=none to window for MPI_Accumulate operations.

bspmm's People

Contributors

rzambre avatar sb17v avatar

Stargazers

 avatar  avatar  avatar  avatar  avatar  avatar  avatar

Watchers

 avatar  avatar  avatar  avatar

Recommend Projects

  • React photo React

    A declarative, efficient, and flexible JavaScript library for building user interfaces.

  • Vue.js photo Vue.js

    ๐Ÿ–– Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.

  • Typescript photo Typescript

    TypeScript is a superset of JavaScript that compiles to clean JavaScript output.

  • TensorFlow photo TensorFlow

    An Open Source Machine Learning Framework for Everyone

  • Django photo Django

    The Web framework for perfectionists with deadlines.

  • D3 photo D3

    Bring data to life with SVG, Canvas and HTML. ๐Ÿ“Š๐Ÿ“ˆ๐ŸŽ‰

Recommend Topics

  • javascript

    JavaScript (JS) is a lightweight interpreted programming language with first-class functions.

  • web

    Some thing interesting about web. New door for the world.

  • server

    A server is a program made to process requests and deliver data to clients.

  • Machine learning

    Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.

  • Game

    Some thing interesting about game, make everyone happy.

Recommend Org

  • Facebook photo Facebook

    We are working to build community through open source technology. NB: members must have two-factor auth.

  • Microsoft photo Microsoft

    Open source projects and samples from Microsoft.

  • Google photo Google

    Google โค๏ธ Open Source for everyone.

  • D3 photo D3

    Data-Driven Documents codes.