Important Information

  • Document: Babel is the cluster hosted in LTI, CMU. Besides this page, please also check the official document. You will need a CMU identity to access this document (i.e., andrew ID).
  • Slack Channel: Babel users should join the babel-babble channel in LTI slack space to receive the latest information. You may also contact the cluster admin through that channel.
  • Use Policy:
    • Generally, each user can use up to 8 GPUs without notifying the admin of the cluster.
    • Occasionally, one can use more than 8 GPUs but need to send a message in the slack channel to clarify the number of GPUs and the estimated time to finish. The admin will request you to lower your usage when the cluster is busy.
    • There is no charging mechanism in babel but please still use it reasonably.
  • swl_general and swl_short partitions:
    • Nodes with names babel-11-* are former SWL cluster. Our lab members will have priority to these nodes as long as you use partitions swl_general and swl_short.

Cluster Access

  • Before you proceed, please make sure your access to Babel is approved by Prof. Shinji Watanabe.
    • Go to LTI intranet and then submit HPC Cluster User Account Request Form.
    • HPC Cluster Name: babel
    • Department Association: LTI
    • Faculty Sponsoring Account: swatanab
    • Additional Groups: swl
  • Connect to the cluster by ssh <username>@babel.lti.cs.cmu.edu

Login nodes, working nodes and working directories

  • Once login, you will be in a login node. These nodes are used for login only and are not for real jobs.
  • You jobs will be conducted by working nodes. You can allocate CPU/GPU resources for your jobs. Once allocated, you can also login these nodes from the login node by ssh. E.g., if there is a job running on babel-11-29, you can login that node by ssh babel-11-29.
  • Working directories below are commonly used. Note /data is not visible to the login nodes.
    • Personal directory: /data/user_data/<user_name>
    • Shared corpus storage /data/group_data/swl/corpora
    • Legacy working directory of previous SWL user: /data/group_data/swl/old_home
    • Personal home, with very limited space. Do not use it for your works: /home/<user_name>

Resource Allocation

  • Resources in Babel are managed by slurm. For general use cases, please refer to this document
  • For ESPnet users, jobs are submitted to slurm automatically.
    • For each recipe (e.g., espnet/egs2/librispeech/asr1), there are a cmd.sh and a conf/slurm.conf files. Setting backend=slurm in cmd.sh and setting conf/slurm.conf properly should be sufficient to use Babel resources. An example conf/slurm.con is below.
      # 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 swl_general --mem 2000M
      option gpu=1 -p swl_general --gres=gpu:1 -c 8  --mem 30000M
      option gpu=2 -p swl_general --gres=gpu:2 -c 16 --mem 60000M
      option gpu=3 -p swl_general --gres=gpu:3 -c 24 --mem 90000M 
      option gpu=4 -p swl_general --gres=gpu:4 -c 32 --mem 120000M 
      option gpu=8 -p swl_general --gres=gpu:8 -c 48 --mem 240000M
      
      • Based on the number of GPUs you request, it will automatically select the setup above. E.g., if 2 GPUs are requested, configuration gpu=2 -p swl_general --gres=gpu:2 -c 16 --mem 60000M will be in use.
      • -p swl_general specify which partition the jobs are submitted to. Use sinfo to check all available partitions. Each partition will contain different resources. Members from WavLab will be able to use partitions debug, general, long, cpu, swl_general and swl_short.
      • -c means the CPU cores to allocate, usually 8 CPU cores for each GPU.
      • --mem means the CPU memory to allocate, usually 30G for each GPU.
      • Make sure gpu=N matches --gres=gpu:N
      • default time=2-00:00:00 specify the estimated time of your jobs. The maximum valid time will be differnt based on the partition. Use sinfo to check that for each partition.
      • Your jobs will fail if the requested number of GPUs / CPU cores / memory beyond the possible configuration.
      • By adding --exclude=<node>, you can avoid submitting your jobs to certain nodes. E.g., --exclude=babel-11-[13,29].
      • By adding -w <nodes>, you can submit your jobs to certain nodes, E.g., --w babel-11-[13,29].
      • You can also specify the GPU types. E.g., to request A6000 GPUs, replace --gres=gpu:4 to --gres=gpu:A6000:4.

ESPnet

Using ESPnet on Babel will not cause extra difficulties. To setup the environment:

git clone https://github.com/espnet/espnet.git
cd espnet/tools
./setup_anaconda.sh <path-to-conda> <env_name> <python_version> # E.g., ./setup_anaconda.sh /data/user_data/<user_name>/tools/miniconda3 espnet 3.10
make TH_VERSION=<torch_version> CUDA_VERSION=<cuda_version> # E.g., make TH_VERSION=2.1.0 CUDA_VERSION=11.8
  • Note: You will not need to use module load as before, as the conda will handle the CUDA automatically.

Then you can run ESPnet recipes. E.g.,

cd espnet/egs2/librispeech/asr1/
# configurate cmd.sh to use slurm backend
# configurate conf/slurm.conf as above
# Add your dataset path to db.sh
bash run.sh

Further ESPnet use guidance is beyond the scope of Babel. Readers can refer to the tutorials in our website.

Misc.

  • VSCode: Both login nodes and working nodes can be accessed by VSCode. Search VSCode in Babel official document for guidance.
  • As /data directory is not visible to login nodes, one can keep a small CPU job for coding. Please only use a small amount of memory / CPU cores for this porpose. For short-time use, you can also allocate some GPUs, but please don’t allocate GPUs for a long time for coding and debugging.

      sbatch --partition=swl_general --nodes=1 --tasks=1 --tasks-per-node=1 --cpus-per-task=4 --mem=8000M -w babel-11-17 --time=15-00:00:00 /home/<user_name>/run.sh & 
    
      ### with the run.sh example below
      #!/bin/bash
      sleep 15d