Account creation

  • Create an account for NCSA on ACCESS
  • Send the username to allocation managers (e.g. Xuankai) to add the user in our group.

login

  • Delta: check https://wiki.ncsa.illinois.edu/display/DSC/Delta+User+Guide#DeltaUserGuide-DirectAccesslogin_nodes

Important

  1. Home directory is of limited space. Please do most of your work in ocean storage ($ cd ${PROJECT})
  2. When you publish a paper, please acknowledge the Delta and ACCESS. We will get benefit when we apply for Delta credits next time.
    • Acknowledgement webpage
      • Example: Experiments of this work used the Bridges2 system at PSC and Delta system at NCSA through allocations CIS210014 and IRI120008P from the Advanced Cyberinfrastructure Coordination Ecosystem: Services \& Support (ACCESS) program, supported by National Science Foundation grants #2138259,#tel:2138286, #tel:2138307, #tel:2137603, and #tel:2138296.

Summary of NCSA usage and the partitions

  • NCSA has limited service units (SUs) for resource availability.
  • General user guide for Delta
  • sinfo lists all the available partitions in NCSA and their status.
NCSA (Delta)
Partitions cpu, gpuA100x4, gpuA100x8, gpuA40x4, gpuMI100x8
GPU resources gpuA100x4, gpuA100x8, gpuA40x4, gpuMI100x8
CPU resources cpu
Default cpu

Other usage

  • Submitting jobs with dependency
    • This can be used to submit a job which is expected to start run after some specific jobs finish. In many cases, training a model can take a few days. However, Delta has the restriction that each job can run for 2 days at most. In this case, we can start a job with dependency for long jobs. For example, you already start a job ID is 000001 and you want a following job right after it. You can submit jobs like: 
        sbatch --time 2-0:00:00 --dependency=afterany:000001 run.sh
  • Common arguments in sbatch / srun 
    -p, --partition=partition               partition requested
-J, --job-name=jobname                  name of job
-t, --time=time                         time limit
--gres=rsrc_name[:rsrc_type]:rsrc_num   required generic resources
-c, --cpus-per-task=ncpus               number of cpus required per task
-d, --dependency=type:jobid[:time]      defer job until condition on jobid is satisfied
-e, --error=err                         file for batch script's standard error
-o, --output=out                        file for batch script's standard output
--mem-per-cpu=MB                        maximum amount of real memory per allocated
--ntasks-per-node=n                     number of tasks to invoke on each node
--reservation=name                      allocate resources from named reservation, e.g. `GPUcis210027`
  • View tools
    • slurm commands 
      # view jobs in the queue
squeue -u ${username}
     
# view detailed job info
scontrol show jobid -d ${jobid}
     
# view job history and billing info, e.g. since time 04/22/2022 12am.
sacct -u ${username} -S 2022-04-22T00:00:00 --format=JobID,jobname,user,elapsed,nnodes,alloccpus,state,partition,nodelist,AllocTRES%50,CPUTime
    • Delta provides slurm-tool 
      # Show or watch job queue:
slurm-tool [watch] queue     show own jobs
slurm-tool [watch] q <user>  show user's jobs
slurm-tool [watch] quick     show quick overview of own jobs
slurm-tool [watch] shorter   sort and compact entire queue by job size
slurm-tool [watch] short     sort and compact entire queue by priority
slurm-tool [watch] full      show everything
slurm-tool [w] [q|qq|ss|s|f] shorthands for above!

slurm-tool qos               show job service classes
slurm-tool top [queue|all]   show summary of active users

# Show detailed information about jobs:
slurm-tool prio [all|short]  show priority components
slurm-tool j|job <jobid>     show everything else
slurm-tool steps <jobid>     show memory usage of running srun job steps

# Show usage and fair-share values from accounting database:
slurm-tool h|history <time>  show jobs finished since, e.g. "1day" (default)
slurm-tool shares

# Show nodes and resources in the cluster:
slurm-tool p|partitions      all partitions
slurm-tool n|nodes           all cluster nodes
slurm-tool c|cpus            total cpu cores in use
slurm-tool cpus <partition>  cores available to partition, allocated and free
slurm-tool cpus jobs         cores/memory reserved by running jobs
slurm-tool cpus queue        cores/memory required by pending jobs
slurm-tool features          List features and GRES

# Other:
slurm-tool v|version         Print versions of slurm-tool tool and slurm itself

ESPnet installation steps

ESPnet installation

You can generally follow the ESPnet installation guide.

Kaldi installation steps (Delta)

   git clone https://github.com/kaldi-asr/kaldi
   cd kaldi/tools
   conda create --name py2 python=2.7 # assume you already have conda environment
   conda activate py2
   conda install -c anaconda mkl mkl-include
   conda install -c conda-forge subversion sox
   make -j 8
   ./extras/install_irstlm.sh # Optional
   cd ../src/
   ./configure --use-cuda=no --mkl-root=<conda_root>/envs/py2
   make -j clean depend; make -j 8

Running ESPnet using slurm backend

Modify cmd.sh and conf/slurm.conf to be ready for slurm management

   # cmd.sh

   # line 31
   cmd_backend='slurm'  # cmd_backend='local'

# conf/slurm.conf

# Default configuration
command sbatch --export=PATH
option name=* --job-name $0
default time=2-00:00:00
option time=* --time $0
option mem=* --mem-per-cpu $0
option mem=0
option num_threads=* --cpus-per-task $0
option num_threads=1 --cpus-per-task 1
option num_nodes=* --nodes $0
default gpu=0
option gpu=0 -p cpu --mem-per-cpu 2000M --account=bbjs-delta-cpu
option gpu=1 -p gpuA40x4 --gres=gpu:1 -c 16 --mem 60000M --account=bbjs-delta-gpu
option gpu=4 -p gpuA40x4 --gres=gpu:4 -c 64 --mem 240000M --account=bbjs-delta-gpu
option gpu=* -p gpuA40x4 --gres=gpu:$0 -c 32 --mem 120000M --account=bbjs-delta-gpu
# Recommend allocating more CPU than, or equal to the number of GPU
# note: the --max-jobs-run option is supported as a special case
# by slurm.pl and you don't have to handle it in the config file.

Running other jobs using slurm backend

  1. Initiate a bash scripts run.sh as
#!/bin/bash

python run.py
  1. Submit the bash script using sbatch
sbatch --time 2-0:00:00 -p gpuA100x4 --gres=gpu:1 -c 16 --mem 60000M --account=bbjs-delta-gpu run.sh

To use a whole node with multiple gpus, you can use

sbatch -n 4 -N 1 --time 48:00:00 -p gpuA40x4 --gres=gpu:4 -c 64 --mem 240000M --account=bbjs-delta-gpu run.sh

ESPnet I/O Issue Fix

Delta will ocassionally have issues with file I/O, randomly preventing files from being accessed programmatically and causing training runs to crash. A simple fix is to catch the I/O exception and re-try. Issues with sound files for example, can be fixed by modifying the sound reader into the following:

# the below code should replace the following: with soundfile.SoundFile(wav) as f: 

try:
  f = soundfile.SoundFile(wav)
except:
  time.sleep(5)
  f = soundfile.SoundFile(wav)