Sbcast Conda Environments
Slurm contains a utility called sbcast that takes a file and broadcasts it to each node’s node-local storage (i.e., /tmp, NVMe).
This is useful for sharing large input files, binaries and shared libraries, while reducing the overhead on shared file systems and overhead at startup.
On Frontier, this is highly recommended at scale if you have multiple shared libraries on Lustre/NFS file systems.
Because Python environments are typically built in this fashion, you may find significant initialization speedup on Frontier if you sbcast your environment to the NVMe (burst buffer) before running any Python scripts.
This guide walks through an example of how to tar up your conda environment using conda-pack and how to sbcast it to the NVMe on Frontier.
OLCF Systems this guide applies to:
Frontier
Installing Conda-Pack
Because conda environments are not relocatable, we must install a tool like conda-pack that will make relocation to the NVMe possible.
Conda-Pack builds archives from original conda package sources and reproduces conda’s own relocation logic.
To install conda-pack, install it from the conda-forge channel like so:
$ conda install -c conda-forge conda-pack
Note
If conda-pack is unable to be installed in your production environment, you can install conda-pack in a separate environment instead and follow a similar workflow.
Installing conda-pack will let you use the conda pack command which can be used to pack your conda environment into a .tar.gz file:
# Pack environment located at an explicit path into my_env.tar.gz
$ conda pack -p /explicit/path/to/my_env
After packing your environment, it can then be moved to the NVMe using sbcast when in a compute job.
Packing your environment will also put a conda-unpack script into the same .tar.gz archive.
Extracting your .tar.gz file and activating your environment will allow you to use the conda-unpack command (script) which will clean up the prefixes of the active environment.
Unpacking your conda environment on the NVMe using conda-unpack will make your conda environment act as if it was installed on the NVMe originally.
The next section will show an example environment on Frontier that is relocated to the NVMe using sbcast.
Example Usage on Frontier
In this example, we will create a new PyTorch environment and move it to the NVMe using conda-pack and sbcast.
First, let’s load our modules and setup the environment:
# Loading the relevant modules
$ module load PrgEnv-gnu/8.5.0
$ module load rocm/5.7.1
$ module load craype-accel-amd-gfx90a
# Create your conda environment
$ module load miniforge3/23.11.0-0
$ conda create -p $MEMBERWORK/<PROJECT_ID>/torch_env python=3.10
$ source activate $MEMBERWORK/<PROJECT_ID>/torch_env
# Install PyTorch w/ ROCm 5.7 support from pre-compiled binary
$ pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm5.7
# Install Conda-Pack into your environment
$ conda install -c conda-forge conda-pack
Next, let’s pack our new conda environment:
$ cd $MEMBERWORK/<PROJECT_ID>
$ conda pack -p $MEMBERWORK/<PROJECT_ID>/torch_env
Finally, let’s run a compute job:
$ sbatch --export=NONE submit.sbatch
Below is an example batch script that uses sbcast, unpacks our environment, and runs an example Python script across 8 nodes:
#!/bin/bash
#SBATCH -A PROJECT_ID
#SBATCH -J bcast_example
#SBATCH -o %x-%j.out
#SBATCH -t 00:05:00
#SBATCH -N 8
#SBATCH -C nvme
date
cd $SLURM_SUBMIT_DIR
# Only necessary if submitting like: sbatch --export=NONE ... (recommended)
# Do NOT include this line when submitting without --export=NONE
unset SLURM_EXPORT_ENV
# Setup modules
module load PrgEnv-gnu/8.5.0
module load rocm/5.7.1
module load miniforge3/23.11.0-0
module load craype-accel-amd-gfx90a
##### START OF SBCAST AND CONDA-UNPACK #####
# Move a copy of the env to the NVMe on each node
echo "copying torch_env to each node in the job"
sbcast -pf ./torch_env.tar.gz /mnt/bb/${USER}/torch_env.tar.gz
if [ ! "$?" == "0" ]; then
# CHECK EXIT CODE. When SBCAST fails, it may leave partial files on the compute nodes, and if you continue to launch srun,
# your application may pick up partially complete shared library files, which would give you confusing errors.
echo "SBCAST failed!"
exit 1
fi
# Untar the environment file (only need 1 task per node to do this)
srun -N8 --ntasks-per-node 1 mkdir /mnt/bb/${USER}/torch_env
echo "untaring torchenv"
srun -N8 --ntasks-per-node 1 tar -xzf /mnt/bb/${USER}/torch_env.tar.gz -C /mnt/bb/${USER}/torch_env
# Unpack the env
source activate /mnt/bb/${USER}/torch_env
srun -N8 --ntasks-per-node 1 conda-unpack
##### END OF SBCAST AND CONDA-UNPACK #####
# Run the Python script
srun --unbuffered -l -N 8 -n 64 -c7 --ntasks-per-node=8 --gpus-per-node=8 --gpus-per-task=1 --gpu-bind=closest python3 example.py
# Gather timings of each slurm jobstep
sacct -j ${SLURM_JOBID} -o jobid%20,Start%20,elapsed%20
The key parts of the above batch script are:
Using the
#SBATCH -C nvmeline makes sure that you’ll get access to the NVMe (accessible at/mnt/bb/<userid>)The
sbcastline broadcasts thetorch_env.tar.gzfile to the NVMe on each nodeYou must make a directory on each NVMe first before extracting the tar file to that directory on each node
Unpacking the environment on each node’s NVMe will make sure each node has access to the new “cleaned” environment
To show the benefit this method provides, let’s see how it affects the timings of running our example script:
import os
import torch
import torch.distributed as dist
def report_env():
rocr_devices = os.getenv("ROCR_VISIBLE_DEVICES")
hip_devices = os.getenv("HIP_VISIBLE_DEVICES")
cuda_visible_devices = os.getenv("CUDA_VISIBLE_DEVICES")
torch_version = torch.__version__
cuda_available = torch.cuda.is_available()
curr_device = torch.cuda.current_device()
device_arch = str(torch.cuda.get_device_name(torch.cuda.current_device()))
cuda_version = torch.version.cuda
hip_version = torch.version.hip
bf16_support = torch.cuda.is_bf16_supported()
nccl_available = torch.distributed.is_nccl_available()
nccl_version = torch.cuda.nccl.version()
print(f"Torch version: {torch_version}")
print(f"CUDA available: {cuda_available} ")
print(f"CUDA version: {cuda_version} ")
print(f"HIP version: {hip_version} ")
print(f"current device: {curr_device} ")
print(f"device arch name: {device_arch} ")
print(f"BF16 support: {bf16_support} ")
print(f"NCCL available: {nccl_available} ")
print(f"NCCL version: {nccl_version} ")
print(f"ROCR_VISIIBLE_DEVICES: {rocr_devices} ")
print(f"HIP_VISIBLE_DEVICES: {hip_devices} ")
print(f"CUDA_VISIBLE_DEVICES: {cuda_visible_devices} ")
def main():
report_env()
if __name__ == "__main__":
main()
Here are the timings from the sbcast NVMe run:
JobID Start Elapsed
--------------- ---------------- --------------------
jobid . 00:01:13
jobid.batch . 00:01:13
jobid.extern . 00:01:13
jobid.0 . 00:00:01 mkdir
jobid.1 . 00:00:49 untar
jobid.2 . 00:00:00 unpack
jobid.3 . 00:00:02 example.py
Here are the timings if the environment was never broadcast from Orion:
JobID Start Elapsed
--------------- ---------------- --------------------
jobid . 00:00:57
jobid.batch . 00:00:57
jobid.extern . 00:00:57
jobid.0 . 00:00:51 example.py
Here are the timings if the environment was stored on NFS and never broadcast:
JobID Start Elapsed
--------------- ---------------- --------------------
jobid . 00:04:04
jobid.batch . 00:04:04
jobid.extern . 00:04:04
jobid.0 . 00:03:56 example.py
The big takeaway is the execution time of example.py, showing that NVMe > Orion >> NFS when it comes to where your conda environment is located before running the script.
Recall, this example was just at 8 nodes and would likely provide more benefit as the node count increases and when using more complex environments (and scripts).
Although extracting the tar.gz file introduces some overhead in the sbcast method, that overhead is small compared to the script initialization overhead in the Orion and NFS method when scaling up to higher node counts.
For more information on using sbcast on Frontier, please see the Frontier User Guide.