Outline

• Efficient file I/O in HPC systems
• Using Allas in batch scripts
• Moving data to/from Allas, IDA and LUMI-O
• Transferring data in sensitive data computing
• Cleaning and backing up data
• Working with remote mounts

Parallel file systems

• A parallel file system (PFS) provides a common file system area that can be accessed from all nodes in a cluster
• Without PFS users would have to always copy all needed data to compute nodes before runs (cf. local disk)
  • Also the results would not be visible outside the compute node
• CSC uses Lustre parallel file system Puhti and Mahti

Lustre

What happens when you access a file?

Managing file I/O (1/3)

• Parallel file system (Lustre):
  • Shared across all nodes in the cluster (e.g., /scratch)
  • Optimized for parallel I/O of large files, slow if accessing lots of small files!
• Temporary local storage (NVMe):
  • Accessible on login nodes ($TMPDIR) and to jobs on some compute nodes ($LOCAL_SCRATCH)
  • Automatically purged after the job finishes
  • Availability varies depending on the supercomputer (Puhti/Mahti/LUMI)
    • For example, Mahti has NVMe only on login nodes and GPU nodes

Managing file I/O (2/3)

• To avoid on Lustre:
  • Accessing lots of small files, opening/closing a single file in a rapid pace
  • Having many files in a single directory
• Use file striping to distribute large files across many OSTs
• Use more efficient file formats when possible
  • Simply using tar and compression is a good start
  • High-level I/O libraries and portable file formats like HDF5 or NetCDF
    • Enable fast I/O through a single file format and parallel operations
    • AI/ML example: TensorFlow’s TFRecords – a simple record-oriented binary format
• Docs CSC: How to achieve better I/O performance on Lustre

Managing file I/O (3/3)

• Use fast local disk to handle file I/O with lots of small files
  • Requires staging and unstaging of data
  • tar xf /scratch/<project>/big_dataset.tar.gz -C $LOCAL_SCRATCH
• Processing data in memory allows better performance compared to writing to and reading from the disk
  • “Ramdisk” (/dev/shm) can be used on Mahti nodes without NVMe
  • export TMPDIR=/dev/shm
• Do not use databases on /scratch
  • Instead, consider hosting DBs on cloud resources (e.g., Pukki DBaaS)

Using Allas in batch jobs

• Command-line interface: use either Swift or S3 protocol
  • Swift (multiple projects, 8-hour) vs. S3 protocol (fixed for a project, persistent)
• allas-conf needs setting up CSC password interactively
  • Jobs may start late and actual job may take longer than 8 hrs
• Use allas-conf -k
  • stores password in variable $OS_PASSWORD to generate a new token automatically
    • a-tools regenerate a token using $OS_PASSWORD automatically
    • rclone requires explicitly setting environment variable in batch jobs:

    source /appl/opt/csc-cli-utils/allas-cli-utils/allas_conf -f -k $OS_PROJECT_NAME

Configuring Allas for S3 protocol

• Opening Allas connection in s3mode
  • source allas_conf --mode s3cmd
• Connection is persistent
• Usage:
  • s3cmd with endpoint s3:
  • rclone with endpoint s3allas:
  • a-put/a-get with -S flag

How to use LUMI-O from Puhti/Mahti?

• LUMI-O is very similar to Allas, but it uses only S3 protocol
• In Puhti and Mahti, connection to LUMI-O can be opened with command:
  • allas-conf --lumi
• Usage:
  • Using LUMI-O with rclone (endpoint is lumi-o:)
    • e.g., rclone lsd lumi-o:
  • One can use a-tools with option --lumi
    • e.g., a-list --lumi
• Docs CSC: Using Allas and LUMI-O from LUMI

Moving data between LUMI-O and Allas

• Requires activating connections to both LUMI-O and Allas at the same time:
  • allas-conf --mode s3cmd
  • allas-conf --lumi
• Use rclone with s3allas: as endpoint for Allas and lumi-o: for LUMI-O
  • rclone copy -P lumi-o:lumi-bucket/object s3allas:allas-bucket/

Moving data between IDA and Allas

• Needs transfer of data via supercomputer (e.g., Puhti)
• Requires configuring IDA in CSC supercomputers.
  • Load IDA module: module load ida
  • Configure IDA database: ida_configure
  • Upload data to IDA: ida upload <target_in_ida> <local_file>
  • Download data from IDA: ida download <target_in_ida> <local_file>

Transferring data for sensitive data computing

• CSC sensitive data services: SD Connect and SD Desktop, use service-specific encryption
• SD Desktop is able to read encrypted data from Allas
  • If you want to make your data available in SD Desktop, you need to use SD Connect to upload data to Allas
  • Use a-put with option --sdx or command a-encrypt to make your Allas data compatible with SD Desktop

Cleaning and backing up data (1/3)

• Disk cleaning
  • In force for project disk areas under /scratch on Puhti
  • Files older than 180 days will be removed periodically
    • Listed in a purge list, e.g. /scratch/purge_lists/project_2001234/path_summary.txt
    • LCleaner tool can help you discover which of your files have been targeted for automatic removal
• Best practice tips
  • Don’t save everything automatically
  • Use LUE tool to analyze your disk usage
    • Avoid du and find -size, these commands are heavy on the file system
  • Move important data not in current use to Allas

Cleaning and backing up data (2/3)

• allas-backup command provides an easy-to-use command-line interface for the restic backup tool
• Backing up differs from normal storing:
  • Incremental (efficient) and version control (no overriding)
  • Based on hashes and requires more computing
  • Efficient way to store different versions of a dataset

Cleaning and backing up data (3/3)

• Please note that Allas is intended for storing active data
• Project lifetime is usually 1-5 years
• Commands for backing up data:
  • allas-backup --help
  • allas-backup [add] file-or-directory
  • allas-backup list
  • allas-backup restore snapshot-id

Working with remote disk mounts

• Using sshfs command in Linux/MacOS:
  • mkdir csc_home
  • sshfs <username>@puhti.csc.fi:/users/<username> csc_home
• To unmount the file system, give the command:
  • fusermount -u csc_home