Support DPDispatcher on Yarn#

Background#

Currently, DPGen(or other DP softwares) supports for HPC systems like Slurm, PBS, LSF and cloud machines. In order to run DPGen jobs on ByteDance internal platform, we need to extend it to support yarn resources. Hadoop Ecosystem is a very commonly used platform to process the big data, and in the process of developing the new interface, we found it can be implemented by only using hadoop opensource components. So for the convenience of the masses, we decided to contribute the codes to opensource community.

Design#

We use DistributedShell and HDFS to implement it. The control flow shows as follows:

Use DistributedShell to submit yarn jobs. It contains generating shell script and submitting it to yarn queues.
Use HDFS to save input files and output results. For performance reasons, we choose to pack forward files to a tar.gz file and upload it to HDFS directory. Accordingly, the task will download the tar file before running and upload result tar file to HDFS after it has done.

Implement#

We only need to add two Class which are HDFSContext and DistributedShell:

class HDFSContext(BaseContext) :
    def upload(self, job_dirs, local_up_files):
    """ upload forward files and forward command files to HDFS root dir

    Parameters
    ----------
    job_dirs : list
        the dictionary which contains the upload files
    local_up_files: list
        the file names which will be uploaded

    Returns
    -------
    none
    """
        pass

    def download(self, submission):
    """ download backward files from HDFS root dir

    Parameters
    ----------
    submission : Submission class instance
        represents a collection of tasks, such as backward file names

    Returns
    -------
    none
    """
        pass

   def check_file_exists(self, fname):
   """ check whether the given file exists, often used in checking whether the belonging job has finished

    Parameters
    ----------
    fname : string
        file name to be checked

    Returns
    -------
    status: boolean
    """
        pass

class DistributedShell(Machine):
    def do_submit(self, job):
    """ submit th job to yarn using distributed shell

    Parameters
    ----------
    job : Job class instance
        job to be submitted

    Returns
    -------
    job_id: string
        usually a yarn application id
    """
        pass

    def check_status(self, job):
    """ check the yarn job status

    Parameters
    ----------
    job : Job class instance
        the submitted job

    Returns
    -------
    status: JobStatus
    """
        pass

    def gen_script_command(self, job):
    """ Generate the shell script to be executed in DistibutedShell container

    Parameters
    ----------
    job : Job class instance
        the submitted job

    Returns
    -------
    script: string
        script command string
    """
        pass

The following is an example of generated shell script. It will be executed in a yarn container:

#!/bin/bash

## set envionment variables
source /opt/intel/oneapi/setvars.sh

## download the tar file from hdfs which contains forward files
if ! ls uuid_upload_*.tgz 1>/dev/null 2>&1; then
    hadoop fs -get /root/uuid/uuid_upload_*.tgz .
fi
for tgz_file in `ls *.tgz`; do tar xvf $tgz_file; done

## check whether the task has finished successfully
hadoop fs -test -e /root/uuid/sys-0001-0015/tag_0_finished
{ if [ ! $? -eq 0 ] ;then
  cur_dir=`pwd`
  cd t sys-0001-0015
  test $? -ne 0 && exit 1

  ## do your job here
  mpirun -n 32 vasp_std  1>> log 2>> err

  if test $? -ne 0; then
      exit 1
  else
      hadoop fs -touchz /root/uuid/sys-0001-0015/tag_0_finished
  fi
  cd $cur_dir
  test $? -ne 0 && exit 1
fi }&

wait

## upload result files to hdfs
tar czf uuid_download.tar.gz sys-0001-0015
hadoop fs -put -f uuid_download.tar.gz /root/uuid/sys-0001-0015

## mark the job has finished
hadoop fs -touchz /root/uuid/uuid_tag_finished

An example of machine.json is as follows, whose batch_type is DistributedShell，and context_type is HDFSContext:

  "fp": [
    {
      "command": "mpirun -n 32 vasp_std",
      "machine": {
        "batch_type": "DistributedShell",
        "context_type": "HDFSContext",
        "local_root": "./",
        "remote_root": "hdfs://path/to/remote/root"
      },
      "resources": {
        "number_node": 1,
        "cpu_per_node": 32,
        "gpu_per_node": 0,
        "queue_name": "queue_name",
        "group_size": 1,
        "source_list": ["/opt/intel/oneapi/setvars.sh"],
        "kwargs": {
          "img_name": "",
          "mem_limit": 32,
          "yarn_path": "/path/to/yarn/jars"
        },
        "envs" : {
          "HADOOP_HOME" : "${HADOOP_HOME:/path/to/hadoop/bin}",
          "CLASSPATH": "`${HADOOP_HOME}/bin/hadoop classpath --glob`",
          "PATH": "${HADOOP_HOME}/bin:${PATH}"}
        }
      }
    }
  ]

Support DPDispatcher on Yarn

Contents

Support DPDispatcher on Yarn#

Background#

Design#

Implement#