Skip to content

Shell scripts for benchmarking Nek5000 across different platforms.

Notifications You must be signed in to change notification settings

thilinarmtb/NekBench

Repository files navigation

NekBench - Scripts for benchmarking Nek5000

(Documentation is outdated, will be updated soon.)

Introduction

This repository contains scripts for benchmarking Nek5000, a fast and scalable open-source spectral element solver for CFD in different platforms from normal linux desktop and laptop machines to super computers at ALCF, NERSC, etc. This is written using bash scripting language and should run without any problem on any Unix-like operating system which supports bash.

Setting up the repository

You can clone the git repository directly using git:

git clone https://github.com/thilinarmtb/NekBench.git

or you can download the repository as a zip archive by clicking the clone or download button in this page.

Running benchmarks

Main script used in benchmarking is the go.sh which can be found on the source root of the repository. If you run ./go.sh --help or ./go.sh -h, it will print out a basic help message describing all the parameters that can be passed into the go.sh script.

go.sh [options]

options:
   -h|--help                 Print this usage information and exit
   -t|--tag name             Specify a tag for the run (unique identifier)
                                 (e.g., scaling, pingpong,...; Default: scaling)
   -m|--machine machine_name Specify a machine for the run
                                 (Mandatory, e.g., linux, theta, cetus, ..)
   -x|--lx1 \"<list>\"         Specify a list of lx1 values for the run
                                 (Mandatory, e.g., \"3 4 5 6\")
   -y|--ly1 \"<list>\"         Specify a list of ly1 values for the run
                                 (Optional, Default: lx1 list)
   -z|--lz1 \"<list>\"         Specify a list of lz1 values for the run
                                 (Optional, Default: lx1 list)
   -e|--lelt \"<list>\"        Specify a list of lelt values for the run
                                 (Mandatory, e.g., \"128 256\")
   -n|--np \"<list>\"          Specify a list of MPI ranks for the run
                                 (e.g., \"2 4 8\"; Default: 1)
   -p|--ppn \"<list>\"         Specify a list of MPI ranks per node for the run
                                 (e.g., \"2 4 8\"; Default: 1)
   -c|--case case_name       Specify the path of the case to be used
                                 in benchmarking (e.g.,/home/nek_user/cases/box)
   --plot plot_type          Plot the benchmark data using matplotlib
                                 (plot_type can be scaling, ping_pong, etc.)
   --even-lxd                Round down lxd to an even value
   clean                     Clean the benchmark directory

Below is an example usage of the go.sh script:

./go.sh -t run1 -m "linux" -x "6" -e "100" -n "2 4 8" -p "2 4 8" -c "/home/foo/NekTests/eddy_uv"

Note that all the parameters can be lists except -t / --tag, -m / --machine and -c / --case. See [1] under section Notes for an explanation why this is.

Once this command is executed, it will create a benchmark run under a folder named benchmarks. The directory structure will look like follows:

├── AUTHORS
├── benchmarks
│   └── run1
│       ├── linux
│       │   ├── lelt_100
│       │   │   ├── lx_6
│       │   │   │   └── eddy_uv
│       │   │   └── lx_8
│       │   │       └── eddy_uv
│       │   ├── lelt_200
│       │   │   ├── lx_6
│       │   │   │   └── eddy_uv
│       │   │   └── lx_8
│       │   │       └── eddy_uv
│       │   └── README
│       └── Nek5000
│           ├── 3rd_party
│           │
.
.
├── bench.sh
├── build.sh
├── functions.sh
├── go.sh
├── jobscripts
│   ├── bgq.submit
│   ├── cori.haswell.submit
.   .
│   └── theta.submit
├── machines
│   ├── bgq
│   ├── cori.haswell
.   .
│   └── theta
├── makeneks
│   ├── bgq.makenek
│   ├── cori.haswell.makenek
.   .
│   └── theta.makenek
├── plot.sh
├── README.md
└── scaling_plots.py

Under the benchmarks directory, there will be a directory named after your tag. Then there will be a directory named after your machine parameter under the tag directory.

Under each machine directory, there will be a separate directory for each value in the lelt list. Similarly, under each of these lelt directories, there will be a separate directory for each lx1 value (Currently, ly1, and lz1 list values are ignored) and the case specified in the script will be copied inside these lower level lx1 directories.

Finally, when the benchmarks are run, this case will be run for all the values in the np list with appropriate ppn value from ppn list if ppn is set. Otherwise, machine specific ppn value set in machines/<machine> will be used.

Currently, NekBench has been succesfully tested on linux laptops/desktops, ALCF Theta, NeRSC cori (KNL and Haswell) and NeRSC Edison machines. Machine (-m / --machine) parameter for each of the previous machines are linux, theta, cori.knl / cori.haswell and edison respectively.

Important notes (must read before using the script !)

  • Case directory name and the case name inside it should be the same i.e., if your case directory is /home/foo/eddy, then there should be eddy.usr, eddy.rea and eddy.map inside that directory.
  • When the go.sh script is run, makenek file for the given machine is searched inside makeneks directory and the job submission script is searched inside jobscripts directory.
  • For example, if you specified -m "linux", makeneks/makenek.linux is used for building the given case and jobscripts/submit.linux is used for running (or submitting, if the machine uses job submission system) the case.
  • Normally, you don't have to change these files. But if you need specific flags in the makenek file, feel free to edit.
  • Each submit.<machine> script takes three arguments. These are fed automatically by the go.sh script. These arguments are different for each machine and depends on the job submission sustem each machine uses (I will add more documentation on these submission files). For each machine, these arguments are created using the file in machines/<machine>. For example, for linux machines, these arguments are created inside machines/linux.

Developer documentation

  • Names of the constants used in the scripts starts with NB_ and variable names start with nb_. If you add new constants/variables, try to follow this standard.
  • functions.sh - is the place to put functions if you end up reusing them again and again.
  • build.sh - creates the directory structure for a particular benchmark run depending on the parameters given to go.sh script.
  • scaling.sh / pingpong.sh - contains logic for scaling and ping-pong benchmarks.

Caveats

  • lx1, lelt, lp, lpmin, lpmax and lxd in your case's SIZE file should have integer expressions initializing them i.e., you can't have something like parameter(lelt=lelg/lpmin + 4), If this is the case, sed substitution fails (We will support these expressions in future).

Notes

  1. Since, tag value is used as the unique identifier for a run within a machine, it does not make sense to have multiple tag values. Also, NekBench is run only in a single machine so it does not make sense to have multiple machine parameters as well.

About

Shell scripts for benchmarking Nek5000 across different platforms.

Topics

Resources

Stars

Watchers

Forks

Packages

No packages published