HPC Resumable Computing Solution

Scenarios

Many HPC applications support resumable computing, such as LAMMPS and GROMACS. In addition, common HPC scheduling software can have resumable computing integrated, such as PBS, Slurm, and LSF.

This section uses LAMMPS as an example to describe how to perform HPC resumable computing.

Step 1: Install FFTW

Run the following commands to install FFTW:

yum install gcc-gfortran gcc-c++

wget http://www.fftw.org/fftw-3.3.8.tar.gz

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/mpi/gcc/openmpi-2.1.2a1/lib64/

export PATH=/usr/mpi/gcc/openmpi-2.1.2a1/bin:$PATH

tar -zxvf fftw-3.3.8.tar.gz

cd fftw-3.3.8/

./configure --prefix=/opt/fftw CC=gcc MPICC=mpicc --enable-mpi --enable-openmp --enable-threads --enable-avx --enable-shared

make && make install

Step 2: Install LAMMPS

1. Run the following commands to install LAMMPS:

   yum install libjpeg-*

   yum install libpng12-*

   wget https://lammps.sandia.gov/tars/lammps-2Aug18.tar.gz

   tar -zxvf lammps-2Aug18.tar.gz

   cd lammps-2Aug18/src

   vi MAKE/Makefile.mpi

2. Modify the data marked in red boxes in Figure 1 and Figure 2. Change the version based on site requirements.
   

   Modify only the data marked in red boxes in Figure 1 and Figure 2.

   Figure 1 Modifying the Makefile file 1
   
   Figure 2 Modifying the Makefile file 2
   
3. Run the following command to compile LAMMPS and copy the obtained lmp_mpi file to /share:

   make mpi

Step 3: Configure LAMMPS

1. Configure the example input file.

   Melt is used as an example to generate example file melt.in. For example, a checkpoint file is automatically generated for every 100 iterative operations, and the file is stored in /share. The file is as follows:

   # 3d Lennard-Jones melt
    
   units           lj
   atom_style      atomic
    
   lattice         fcc 0.8442
   region          box block 0 20 0 20 0 20
   create_box      1 box
   create_atoms    1 box
   mass            1 1.0
    
   velocity        all create 1.44 87287 loop geom
    
   pair_style      lj/cut 2.5
   pair_coeff      1 1 1.0 1.0 2.5
    
   neighbor        0.3 bin
   neigh_modify    delay 5 every 1
    
   fix             1 all nve
   dump 1 all xyz 100 /share/sample.xyz
   run             10000 every 100 "write_restart /share/lennard.restart"

2. Obtain the melt.restart.in input file for resumable checkpoint computing.

   # 3d Lennard-Jones melt
    
   read_restart  /share/lennard.restart
   run        10000 every 100 "write_restart /share/lennard.restart"

3. Obtain the PBS job script job.pbs.

   #!/bin/sh
   #PBS -l ncpus=2
   #PBS -o lammps_pbs.log
   #PBS -j oe
    
   export PATH=/usr/mpi/gcc/openmpi-2.1.2a1/bin:$PATH
    
   export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/mpi/gcc/openmpi-2.1.2a1/lib64/module 
    
   if [ ! -e "/share/lennard.restart" ]; then
     echo "run at the beginning"
     mpiexec --allow-run-as-root -np 2 /share/lmp_mpi -in /share/melt.in
   else
     echo "run from the last checkpoint"
     mpiexec --allow-run-as-root -np 2 /share/lmp_mpi -in /share/melt.restart.in
   fi

Step 4: Submit a Job and Ensure that the Job Is Not Interrupted During Running

Submit and run the job without interrupting it, and check the job running duration.

1. Run the following command to submit a job:

   qsub job.pbs

2. After the job is complete, run the following command to view the job information:

   qstat -f Job ID

   As shown in Figure 3, the job runs for 4 minutes and 10 seconds.

   Figure 3 Job running without being interrupted
   

Step 5: Submit a Job, Emulate a Computing Interrupt, and Use Resumable Computing to Complete the Computing

After submitting a job, stop the compute node to emulate a computing interruption. Then, check the job running durations before and after the interruption.

1. Run the following command to submit a job:

   qsub job.pbs

2. After the job runs for about 1 minute and 30 seconds, stop the compute node on which the job runs to emulate example release.
3. Run the following command to check the job information after the compute node is stopped:

   qstat -f Job ID

   Figure 4 Job running before interruption
   

   In such a case, the job returns back to the queued state, waiting for available computing resources.

4. Start the compute node to provide available computing resources.

   In such a case, the job will continue.

5. After the job is complete, run the following command to view the job information:

   qstat -f Job ID

   As shown in Figure 5, the job runs for 3 minutes and 3 seconds. It is shown that the job computing is resumed at the time when the computing is interrupted.

   Figure 5 Job running after interruption