Stage Data for Compute Jobs
Large datasets are often stored off-cluster on departmental servers, Storage@Yale, in cloud storage, etc. Since the permanent home of the data remains on off-cluster storage, you need to transfer a working copy to the cluster temporarily. When your computation finishes, you would then remove the copy and transfer the results to a more permanent location.
Temporary Storage
We recommend staging data into your scratch storage space on the cluster, as the working copy of the data can then be removed manually or left to be deleted (which will happen automatically after 60-days).
Interactive Transfers
For interactive transfers, please see our Transfer Data page for a more complete list of ways to move data efficiently to and from the clusters.
A sample workflow using rsync would be:
# connect to the transfer node from the login node
[netID@cluster ~] ssh transfer
# copy data to temporary cluster storage
[netID@transfer ~]$ rsync -avP netID@department_server:/path/to/data $HOME/palmer_scratch/
# process data on cluster
[netID@transfer ~]$ sbatch data_processing.sh
# return results to permanent storage for safe-keeping
[netID@transfer ~]$ rsync -avP $HOME/palmer_scratch/output_data netID@department_server:/path/to/outputs/
Tip
To protect your transfer from network interruptions between your computer and the transfer node, launch your rsync inside a tmux session on the transfer node.
Transfer Partition
Both Grace and McCleary have dedicated data transfer partitions (named transfer) designed for staging data onto the cluster.
All users are able to submit jobs to these partitions. Note each users is limited to running two transfer jobs at one time.
If your workflow requires more simultaneuous transfers, contact us for assistance.
Transfers as Batch Jobs
A sample sbatch script for an rsync transfer is show here:
#!/bin/bash
#SBATCH --partition=transfer
#SBATCH --time=6:00:00
#SBATCH --cpus-per-task=1
#SBATCH --job-name=my_transfer
#SBATCH --output=transfer.txt
rsync -av netID@department_server:/path/to/data $HOME/palmer_scratch/
remote.host.yale.edu to your scratch directory.
Note, this will only work if you have set up password-less logins on the remote host.
Transfer Job Dependencies
There are sbatch options that allow you to hold a job from running until a previous job finishes.
These are called Job Dependencies, and they allow you to include a data-staging step as part of your data processing pipe-line.
Consider a workflow where we would like to process data located on a remote server. We can break this into two separate Slurm jobs: a transfer job followed by a processing job.
transfer.sbatch
#!/bin/bash
#SBATCH --partition=transfer
#SBATCH --time=6:00:00
#SBATCH --cpus-per-task=1
#SBATCH --job-name=my_transfer
rsync -av netID@department_server:/path/to/data $HOME/palmer_scratch/
process.sbatch
#!/bin/bash
#SBATCH --partition=day
#SBATCH --time=6:00:00
#SBATCH --cpus-per-task=1
#SBATCH --job-name=my_process
module purge
module load miniconda
conda activate my_env
python $HOME/process_script.py $HOME/palmer_scratch/data
First we would submit the transfer job to Slurm:
$ sbatch transfer.sbatch
Submitted batch job 12345678
Then we can pass this jobID as a dependency for the processing job:
$ sbatch --dependency=afterok:12345678 process.sbatch
Submitted batch job 12345679
$ squeue
JOBID PARTITION NAME USER ST TIME NODES NODELIST(REASON)
12345679 day process netID PD 0:00 1 (Dependency)
12345678 transfer transfer netID R 0:15 1 c01n04
Storage@Yale Transfers
Storage@Yale shares are mounted on the transfer partition, enabling you to stage data from these remote servers.
The process is somewhat simpler than the above example because we do not need to rsync the data, and can instead use cp directly.
Here, we have modified the transfer.sbatch file from above:
transfer.sbatch
#!/bin/bash
#SBATCH --partition=transfer
#SBATCH --time=6:00:00
#SBATCH --cpus-per-task=1
#SBATCH --job-name=my_transfer
cp /SAY/standard/my_say_share/data $HOME/palmer_scratch/
This will transfer data from the Storage@Yale share to palmer_scratch where it can be processed on any of the compute nodes.