ioComp

Library to compare performance from different I/O libraries with the option to switch between asynchronous and direct synchronous I/O. Different back-ends include MPIIO, HDF5, ADIOS2 HDF5 and ADIOS2 BP4, BP5

Organization

The repository is arranged as follows:

<root>
    include/                       # Header files for iocomp library 
    src/                           # iocomp library source code 
    stream/                        # STREAM test benchmark
        run_dir/                   # Directory containing all slurm scripts to run on ARCHER2 and Cirrus  
    test/                          # Testing source code with read scripts for output formats 
    LICENSE                  	   # The software license governing the software in this repository

Installation for iocomp library

git clone <iocomp repository>
cd iocomp 
cd src 
mkdir Object_files
make DIR=</path/to/installation/dir> HDF5_DIR=<path/to/hdf> ADIOS2_DIR=<path/to/adios2>
# if only MPI compilation required
make DIR=</path/to/installation/dir> NOADIOS2=1 NOHDF5=1

Preprocessor flags for the iocomp src/Makefile

• NDELETE : add this flag to not delete the files after finishing writing
• NDEBUG : add this flag to not print out error messages
• IOCOMP_TIMERS : add this flag to print out I/O bandwidths after each step of writing is completed
• PRINT_ORDERING : add this flag to print out ordering between the pairs
• READBACK : add this flag to print out the file that was just written by the selected I/O back-end

Installation for iocomp STREAM implementation

To test the library with a STREAM implementation:

cd iocomp 
cd stream 
mkdir Object_files
make 

iocomp run commands

Command line arguments

• nx : Dimension 1 of STREAM vector.
• ny : Dimension 2 of STREAM vector. nx * ny give the total size of array that is passed to STREAM.
• io : I/O backend selection.
  • 0 : MPIIO
  • 1 : HDF5
  • 2 : ADIOS2_HDF5
  • 3 : ADIOS2_BP4
  • 4 : ADIOS2_BP5
• HT : Flag which when passed splits the communicators to have different operations on different processes.

for the default case with sequential compute and I/O operations

./test --nx=<size of nx> --ny=<size of ny> --io=<I/O selection>

for the HT case with compute and I/O operations on different MPI processes

./test --HT --nx=<size of nx> --ny=<size of ny> --io=<I/O selection>

Run jobs on ARCHER2

runscript.sh contains the following variables:

• IO_START : starting I/O library
• IO_END : ending I/O library
• NX : Dimension 1 of STREAM vector
• NY : Dimension 2 of STREAM vector. nx * ny give the total size of array that is passed to STREAM.
• DIR : parent directory to store output files and sub directories
• QOS : standard or lowpriority determines the quality of service on ARCHER2
• TIME : array of times per job
• NODE_START : starting value for loop over number of nodes as power of 2s.
• NODE_END : ending value for loop over number of nodes as power of 2s.
• ARRAY : SLURM job array options for averaging
• weakScaling.sh : Source weakScaling script to run weak scaling jobs
• strongScaling.sh : Source strongScaling script to run strong scaling jobs

run scripts for ARCHER2

cd run_dir
bash runscript.sh