Hyper IO Testing

Using Hyper IO to Accelerate Application Performance

HyperIO_WRF_1

1. The shared storage is mounted properly on eight Kunpeng 920 servers, whose RoCE communication is normal.
2. openEuler 22.03 has been deployed and is running properly.
3. The default compiler is used.
4. Hyper MPI 1.3.1 has been deployed and Hyper MPI environment variables have been loaded.
5. WRF 4.2.2 has been deployed and the test case is prepared.
6. Hyper IO has been decompressed and deployed.
7. You have logged in to the control terminal of the server.

1. Obtain the test case file.

   mkdir -p /path/to/WRFTEST

   cd /path/to/WRFTEST

   wget https://www2.mmm.ucar.edu/WG2bench/conus_2.5_v3/1-RST/RST/ -r -c -np -nH --cut-dirs 2 --restrict-file-names=nocontrol -e robots=off --reject "index.html*"

   wget https://www2.mmm.ucar.edu/WG2bench/conus_2.5_v3/wrfbdy_d01.gz -r -c -np -nH --cut-dirs 2 --restrict-file-names=nocontrol -e robots=off --reject "index.html*"

   wget https://www2.mmm.ucar.edu/WG2bench/conus_2.5_v3/1-RST/namelist.input -r -c -np -nH --cut-dirs 2 --restrict-file-names=nocontrol -e robots=off --reject "index.html*"

2. Generate a test case.

   cat rst_6hr* | gunzip -c > wrfrst_d01_2005-06-04_06_00_00

3. Copy WRF-4.2.2/run/* to /path/to/WRFTEST:

   cp /path/to/WRF/WRF-4.2.2/run/* ./

   rm -rf wrf.exe

   ln -s /path/to/WRF/WRF-4.2.2/main/wrf.exe wrf.exe

4. Decompress the wrfbdy_d01 file.

   gunzip wrfbdy_d01.gz

5. Modify the namelist.input file.
6. Run the WRF program and record the running duration.

   time mpirun --allow-run-as-root -N 8 -np 640 -x PATH -x LD_LIBRARY_PATH --hostfile hf4 OMP_NUM_THREADS=1 /path/to/WRFTEST/wrf.exe

7. Run the WRF program using Hyper IO and record the running duration.

   time mpirun --allow-run-as-root -N 8 -np 640 -x PATH -x LD_LIBRARY_PATH -x LD_PRELOAD=/path/to/IO_Middleware_version_aarch64/lib/libock_shore_pnetcdf.so -x OCK_IO_CONFIG=/path/to/wrf.json --hostfile hf4 OMP_NUM_THREADS=1 /path/to/WRFTEST/wrf.exe

8. Compare the running durations in 6 and 7.

The running duration of the WRF application is shortened by more than 10%.

-

-