MPI NBC Benchmark

Documentation

-*- mode: Markdown;-*-

README for Bench NBC

This document describes Bench NBC installation and configuration.

For any question, mailto: Florian.Reynier@inria.fr and Alexandre.Denis@inria.fr

For more information, see: http://pm2.gitlabpages.inria.fr/bench_nbc/

Quick Start

A quick cheat sheet for the impatient: 

./configure
make
./bench_raw_estim
mpiexec -n 64 ./bench_nbc -C iallgather -b bench_raw_estim.out -T -o out

It runs from 30 minutes through 2h, depanding on network speed and number of nodes. It will create data files based on the "out" token, in particular out_raw.dat. Then build the performance report using: 

./bench_nbc_gen_graphs out_raw.dat

It outputs pdf graphs. It is possible to transfer data to another host and extract the performance report with another installation of Bench NBC so as to not have to install R on the computing nodes.

Please send the out_raw.dat to Flori.nosp@m.an.R.nosp@m.eynie.nosp@m.r@in.nosp@m.ria.f.nosp@m.r and Alexa.nosp@m.ndre.nosp@m..Deni.nosp@m.s@in.nosp@m.ria.f.nosp@m.r to have it integrated on the Bench NBC benchmark web site.

Requirements

  • MPI library
  • autoconf (v 2.50 or later, for git users)
  • OpenMP compiler
  • hwloc
  • R (optional, only to generate graphs)
  • doxygen (optional, for doc generation)

Installation

Bench NBC benchmark follows usual autoconf procedure: 

./configure [your options here]
make
make install

The make install step is optional. The benchmark may be run from its build directory. To get help on supported flags for configure, run: 

./configure --help

Flags that may be of interest are MPICC= to give the name of the command to build MPI applications, and –prefix= to give installation path.

Documentation

  • Bench NBC needs a first non-MPI calibration, performed by bench_raw_estim. You can give an output filename on the command line to bench_raw_estim. If no filename is given, the default output is bench_raw_estim.out.

  • In a second pass, run the MPI benchmark bench_nbc. It needs the following args:

    -C <op> the collective operation to benchmark. Supported operations are:

    • ibarrier, ibcast, ireduce, ialltoall, iallreduce, igather, iallgather

    -b <file> the filename containing the output from the first non-MPI benchmark.

    -o <name> a basename used to output data

    -T use steps by time instead of data size for collective. This parameter is mandatory with the current version.

  • To generate graphs from data, a set of R scripts are available in the script/ directory. For convenience, command bench_nbc_gen_graphs calls them in sequence so as to generate all graphs at once.