Getting Started with slk#

file version: 09 Jul 2024

current software versions: slk version 3.3.91; slk_helpers version 1.12.10; slk wrappers version 1.2.2

Introduction#

slk and slk_helpers are the command line interfaces that have to be used to interact with the HSM system StrongLink (tape archive) at DKRZ. slk can be used to archive data to tape and to list, search and retrieve archived data on/from tape. It is provided by StrongLink. The slk_helpers provide additional functionality and are developed and maintained at DKRZ. Before you can use the slk and/or slk_helpers commands the first time, you must call slk login and login with your DKRZ/luv login credentials.

The slk and slk_helpers commands behave very similar to known commands on the Linux terminal. Not all slk and slk_helpers commands are listed below. Please have a look into the slk manual and slk_helpers manual to get lists of all available commands.

  • cp => slk archive and slk retrieve

  • ls => slk list and slk_helpers list_search

  • rm => slk delete

  • mv => slk move and slk rename

  • chmod => slk chmod

  • chgrp => slk group

  • chown => slk owner (only admins)

  • find => slk search

  • mkdir => slk_helpers mkdir (slk archive automatically creates namespaces)

  • sha512sum => slk_helpers checksum -t sha512

  • test -e => slk_helpers exists (also returns the resource id)

  • get file size, e.g. equal to stat --printf="%s" => slk_helpers size

folder or directory is called namespace in StrongLink.

Please do not run slk archival and slk retrieval on Levante login nodes if your plan to transfer more than a few GB. Instead, please run them on the shared or interactive nodes as described here and here, respectively, and request 6GB of memory (--mem=6GB). If your slk is killed with a message like /sw/[...]/bin/slk: line 16: [...] Killed, then then your slk was killed due to too less available memory.

Some slk commands do not print textual output to the SLURM log when used in batch jobs. Therefore, we strongly recommend capturing the exit code $? of slk as described in the slk in SLURM jobs. In SLURM jobs, slk archive prints its double verbose output (-vv) to the SLURM log but not its single verbose output (-v). Additionally, slk archive, slk retrieve and slk recall print a summary report to the slk log file (~/.slk/slk-cli.log) which contains information on whether the commands finished successfully or not.

Start using slk and slk_helpers#

Load slk module#

The slk and slk_helpers are available as module on all Levante nodes. Just do

$ module load slk

to load the most recent program versions. In the past, slk and slk_helpers were available in separate modules. These modules have been merged. The new module contains three components:

  • slk

  • slk_helpers

  • sbatch wrapper scripts for slk and slk_helpers

If you wish to use a specific old release (e.g. slk_helpers version 1.10.2), please do

$ module load slk/3.3.91_h1.10.2_w1.2.1

This will load slk version 3.3.91, slk_helpers version 1.10.2 and the sbatch wrappers version 1.2.1.

Please be aware that old slk and slk_helpers releases are not officially supported anymore, contain more bugs and might not be 100% compatible with the current version of StrongLink.

Login#

Call slk login and use your common DKRZ credentials to log in. You do not login to another shell with slk login (like pftp) but it creates a login token. This login token allows the usage of the other slk commands. It is located in ~/.slk/config.json and is valid for 30 days.

$ slk login

All Levante nodes share the same home directory. Therefore, you need to login only once every 30 days and you can run it on any Levante node.

Reminder login token expires#

If you wish to be reminded when the login token is due to expire, you can set up SLURM job to check the expiration data of the login token on daily basis. You might use this script for this purpose. It sends an email to you when the token expires in 30, 5, 4, 3, 2 and 1 days. The email is send to email address which is stored in your user profile at https://luv.dkrz.de . The whole output is written into the file slk_token_expire_check.log which is automatically kept below a size of 1100 lines. When finished, the script submits itself to start again the next day at the same time. Please submit the script this way replacing YOUR_PROJECT_ACCOUNT by a valid project/account with compute time:

sbatch --begin="now+1day" --account=YOUR_PROJECT_ACCOUNT ./slk_token_expire_email.sh

slk in SLURM jobs#

Please capture the exit code ($?) of each slk command which is run in a SLURM/batch job because some slk commands do not print textual output when used in batch jobs (see How do I capture exit codes?; example: example batch scripts).

Please run slk archive and slk retrieve with 6 GB of memory (--mem=6GB in batch jobs). Detailed diagnostic information is printed into the slk log file: ~/.slk/slk-cli.log. The command line output might be misleading in some situations as documented in the Known Issues. If your slk is killed with a message like /sw/[...]/bin/slk: line 16: [...] Killed, then please inform the DKRZ support (support@dkrz.de) and re-run the slk command with 8 GB or 10 GB of memory.

slk memory footprint#

slk archive and slk retrieve might use up to 4 GB of memory and cause high CPU load when large files or many files are archived/retrieved. Running these tasks on the login nodes might affect other users working on these nodes. Hence, we encourage all users to perform these tasks on nodes of the compute, interactive or shared partitons and with sufficient amount of memory allocated. In order to account for overhead and similar, 6 GB of memory should be allocated in the job script via --mem=6GB (examples: Batch Script Examples; --mem in sbatch Manual: https://slurm.schedmd.com/sbatch.html#OPT_mem).

slk data transfer rate#

Transferring files with a single call of slk archive or slk retrieve reaches transfer rates of up to 1 GB/s and might even slightly exceed this rate under favorable conditions. The transfer rate can be lower if the StrongLink constellation is under high load. Running _several_ slk archive or slk retrieve in parallel on one node will not necessarily increase the total transfer rate because at some point the transfer rate is limited by the hardware. We are still lacking experience on how many parallel slk archive / retrieve are still reasonable. Generally, please avoid running too many individual parallel calls of slk retrieve but try to bundle the retrieval of individual files as suggested in search and retrieval of search results. If you run several calls of slk in parallel, please make sure to allocate sufficient memory.

Archival#

The slk archive command is available on all nodes of Levante and can be used to archive files or directories. Similar to cp it prints not output by default. A progress bar is printed when -v is set in interactive mode. A detailed list of all processed files is printed when -vv is set (also in batch mode). slk archive allows the usage of wildcards *, ? and [...]. Add -R to do recursive archival of directories. Basic examples slk archive calls are:

# archive one file, absolute path
$ slk archive /work/bm0146/k204221/some_file.nc /arch/ab01234/c567890/my_data_1/
# archive one file, relative path
$ slk archive some_file.nc /arch/ab01234/c567890/my_data_3/
# archive folder recursively, absolute path
$ slk archive -R /work/bm0146/k204221/some_folder /arch/ab01234/c567890/my_data_4/
# archive folder recursively, relative path, skip hidden files and folders
$ slk archive -x -R some_folder /arch/ab01234/c567890/my_data_5/
# archive multiple files
$ slk archive some_file_a.nc some_file_b.nc some_file_c.nc /arch/ab01234/c567890/my_data_6/
# archive multiple files using wildcards
$ slk archive file_?.nc /arch/ab01234/c567890/my_data_7/
$ slk archive year200[0123].nc /arch/ab01234/c567890/my_data_8/

Please run it via salloc as interactive job in partition interactive or in a batch script in partition shared if you plan to archive a file larger than a few GB. In both cases, 6 GB of memory should be allocated (--mem=6GB).

The optimal file size is between 10 GB and 200 GB. Each file takes up a least 1 GB of archival quota. Retrieving files smaller than 1 GB is very inefficient and, hence, we strongly recommend not to archive smaller files. Instead, we recommend to pack small files to tar-balls of the optimal size. packems is very useful for this purpose. Hidden files and folders, such as .git or .config, are ignored when -x is set. Please avoid archiving more than approximately 3 TB with one call of slk archive. Archiving higher amounts of data at once might cause the slk archive to fail (details: How much data can I archive at once?).

When you run slk archive with -vv it will print the archival status of each file it tried to archive.

The group of an archived file is always the default group of the archiving user and not the group of the project folder into which the file is archived. The owner has to change the group manually after archival.

Please check if slk archive finishes successfully. This can be done by one of three means:

  • look into the textual output of slk archive,

  • capture the exit code $? (0: successful archival)

  • look into the slk log file: ~./slk/slk-cli.log

If the transfer via slk archive is interrupted, an incomplete version of this file remains in the HSM. If you re-run the same slk archive command it transfers only incomplete (overwrites them) and missing files. Existing files, which were archived successully already, are skipped. Thus, slk archive works like rsync with this respect.

Incomplete files are listed by slk list and are not deleted automatically by the StrongLink system. Please do not trust the partial file info printed by slk list. In order to check whether a file was completely archived or not, please run a verify job. Once or twice a week the DKRZ staff searches for incompletely archived files, informs the file owners and moves these files to a specific location.

Generally, you can verify archived files by running a verify job or by comparing the sha512 checksums.

Pack & Archive#

The package packems, which was developed by MPI-M and DKRZ for the HPSS, has been adapted to the StrongLink system / slk. The process of packing & archiving of multiple data files to tape and their retrieval is simplified by this package. It consists of three command line programs:

  • a pack-&-archive tools packems,

  • a list-archived-content tool listems,

  • a retrieve-&-unpack tool unpackems and

Currently, there is a low probability that slk archive and other slk commands fail with connection timeouts or out of memory errors. When this happens, parts of the packems workflow may fail – and failure might be overlooked. Therefore, we propose to run packems twice: first, to generate tar balls and, second, to archive the tar balls and to generate an index. The second run of packems needs sufficient memory being allocated as mentioned in the comments below. This would be an example workflow:

# set some variables for better understanding
# TODO: please set these three environment variables according to your needs
source_path=/work/ab1234/c567890/scenario_xyz/output
tmp_path=/scratch/c/c567890/packems_tmp
destination_path=/arch/ab1234/model_output
tar_base_name=scenario_xyz

# load packems module
module load packems slk

# pack data with packems (without archival)
# if you are in a SLURM job on a compute node you can increase `-j 2` to higher values -- e.g. `-j 16`
# TODO: please add additional flags to the packems call according to your needs
packems -j 2 ${source_path} -d ${tmp_path} -o ${tar_base_name} --no-archive

# archive packed data with packems (packing already done previously)
# NOTE: Each running archival process needs approx. 6 GB of memory. This means:
#           * do not run packems on login nodes
#           * allocate sufficient memory with `--mem=<memory>`
#           * set `-j <parallel tasks>` accordingly (`<memory> = <parallel tasks> * 6 GB`)
packems -j 1 ${source_path} -d ${tmp_path} -o ${tar_base_name} --archive-only

For details on the usage of packems please have a look into the packems manual.

List content#

slk list will print the content of a namespace similar to ls -la. The file size is automatically printed human-readably. Please set -b to print the file sizes in bytes.

$ slk list /arch/bm0146/k204221/my_data
-rw-rw-rw-- k204221     bm0146          1.2K   27 Mar 2020  13:18  borehole_01.nc
-rw-rw-rw-- k204221     bm0146          1.2K   04 Mar 2021  10:13  nc3.nc
-rw-rw-rw-- k204221     bm0146          1.2K   04 Mar 2021  09:29  nc_k_2.nc
-rw-rw-rw-- k204221     bm0146          4.0M   04 Mar 2021  17:22  nc_k_3.nc
-rw-rw-rw-- k204221     bm0146          4.0M   04 Mar 2021  10:02  nc_k_4.nc
-rw-rw-rw-- k204221     bm0146         13.1K   08 Dec 2020  22:29  small.nc
-rwxrwx-w-t k204221     bm0146        105.5M   08 Nov 2019  09:05  small_BPb4-Sl-mT_00062104_00040000000_01040000000.AGM07807972.freeze.nc
-rwxrwx-w-t k204221     bm0146        105.5M   14 Nov 2019  03:23  small_BPb4-Sl-mT_00062104_00040000000_01040000000.AGM07807972.nc
-rw-rw-rw-- k204221     bm0146          1.2K   23 Mar 2021  22:41  test.nc
-rw-rw-rw-- k204221     bm0146          1.2K   23 Mar 2021  20:04  zonk.nc
-rw-rw-rw-- k204221     bm0146          1.2K   28 Jun 2021  14:26  some_file.nc
Files: 11

Validate archivals#

This content has been moved to section Validate archivals on the page Archivals to tape.

Retrievals#

Warning

If you retrieve more than 10 files are once, please run slk_helpers gfbt PATH -R --count-tapes or slk_helpers gfbt --search-id SEARCH_ID --count-tapes first. If the files are located on more than five tapes, please split the retrieval into multiple parts as described here (see also short command manual and usage examples). Currently, the whole StrongLink system slows down considerably, when single retrievals / recalls access too many tapes at once. Starting many one-file-recalls at once might slow down the system as well.

slk retrieve is the counterpart of slk archive. Please do not use bash globs / wildscards with slk retrieve. Instead, please perform a search and use the resulting search id as input for slk retrieve.

If used on a Levante login node, slk retrieve can only retrieve one file at once. To retrieve more than one file at once, please run slk retrieve on nodes of the compute, shared and interactive partitions of Levante. If you wish to use slk retrieve interactively, please start an interactive batch session via the interactive partition with salloc (Run slk in the “interactive” partition and Data Processing on Levante).

This is an example call of slk retrieve:

slk retrieve /arch/bm0146/k204221/my_data/nc3.nc /work/bm0146/k204221/results

If slk retrieve is run to retrieve a file, it checks whether this file is already stored in the HSM cache or not.

  • If the file is stored in the HSM cache then it is copied directly from the cache to Levante.

  • If the file is not stored in the HSM cache then a recall job is started in StrongLink which copies the file from tape to the HSM cache. When the file has arrived in the HSM cache it is automatically copied from the cache to Levante.

In contrast to slk retrieve, slk recall will copy a file from tape to cache but not to the Lustre filesystem. This command can be used to prestage files and to retrieve them later:

slk recall /arch/bm0146/k204221/my_data/nc3.nc

Each process in StrongLink, which copies data from tape to the HSM cache, has its own job id. This id is printed into the slk log ~/.slk/slk-cli.log by slk retrieve and slk recall. If your retrieval/recall takes very long or your slk command is aborted, you can check the status of the copy job via slk_helpers job_status JOB_ID. If the StrongLink system is very busy, jobs are queued. You can check the queue lenght via slk_helpers job_queue. For details please read Waiting and processing time of retrievals.

You might use the wrapper script slk_wrapper_recall_wait_retrieve to start SLURM jobs, which initiates a recall, waits until the recall job is finished and, then, retrieves the files. This approach simplifies the debugging is an error occurs and saves nodes hours. The script writes all relevant output into one log file which name can be partly set by the user.

slk slk_wrapper_recall_wait_retrieve <ACCOUNT> /arch/bm0146/k204221/my_data/nc3.nc /work/bm0146/k204221/results <SUFFIX FOR THE LOG FILE>

If slk retrieve or slk recall is aborted while files are copied from tape, please check via the job id whether StrongLink is still copying the files back from tape to cache. If so, please do not restart the commands until the job has finished. Restarting retrievals / recalls for files from tape multiple times might cause the StrongLink system to slow down.

When multiple files, which are stored on one tape, should be retrieved, it is inefficient to retrieve each file with a single call of slk retrieve because – in the worst case – the tape might be loaded and ejected for each file individually. Instead, all files, which are located on one tape, should be retrieve with one call of slk retrieve. There are two ways to do this:

  1. Perform a search for all files that should be retrieved and then retrieve the search results as suggested in Run a search query and retrieve search results and this example script. This will decrease the overall time needed for the retrieval because the files are read more efficiently from tapes. In this case, data from more than one tape might be accessed. If one of these tapes is not available in the moment, the whole retrieval might fail. Additionally, the StrongLink system will slow down if you retrieve data from more than 10 tapes with one call of slk retrieve. Please check in advance via slk_helpers gfbt --search-id SEARCH_ID --count-tapes.

  2. The best performance and error control is reached if all needed data which are stored on one tape are retrieved/recalled with one call of slk retrieve / slk recall. At the same time, each slk retrieve call should target only one tape. This can be done via the command slk_helpers group_files_by_tape / slk_helpers gfbt. The command usage of the command is described here (see also short command manual and usage examples).

Note

We recommend using striping on Levante for the time being. Some folders are striped already. Please check in advance.

Move, rename and delete files or namespaces#

Move a file from one namespace to another. The file’s name cannot be changed by this command.

$ slk move /arch/bm0146/k204221/my_data/nc_k_2.nc /arch/bm0146/k204221/old_data

Warning

slk move automatically overwrites target files if they exit (similar to mv). Please run slk move with -i avoid this.

Rename a file. The file’s location cannot be changed by this command.

$ slk rename /arch/bm0146/k204221/my_data/nc_k_3.nc a_netcdf_file.nc

Delete a file. If we apply slk delete onto a namespace then it deletes all files in this namespace without confirmation. Files in sub-namespaces are not deleted. To delete the whole namespace with all sub-, subsub-, …-namespaces and their content, please append -R. You may supply more than one file to slk delete to delete more than one file at once.

$ slk delete /arch/bm0146/k204221/my_data/zonk.nc

Search files#

Files stored in StrongLink can be searched via search queries, which are written in JSON. You can use jq to print the search queries in a human-readable way. Details on the format of search queries are given in Metadata in StrongLink.

The slk provides search functionality via the command slk search. After running a search query, it returns a search ID that can be used with slk list and slk retrieve to list and retrieve, respectively, the found data. Below you will find three example tasks for searches. More examples and explanations are given here: Metadata in StrongLink.

Task: find all *.nc files in the namespace /ex/am/ple/data

slk_helpers search '{"$and": [{"path": {"$gte": "/ex/am/ple/data"}}, {"resources.name": {"$regex": ".nc$"}}]}'

Task: find all netCDF files, which have a global attribute project with the value ABC. For netcdf.Project below please note that netcdf is written lower case and Project starts with an upper case P although the global attribute might have been written with lower case p.

slk_helpers search '{"netcdf.Project": "ABC"}'

Task: find all INDEX.txt files that either belong to user k204221 (uid: 25301) in /arch or that are stored in the namespace /double/bm0146.

slk_helpers search '{"$and": [{"resources.name": "INDEX.txt"}, {"$or": [{"$and": [{"resources.posix_uid": 25301}, {"path": {"$gte": "/arch"}}]}, {"path": {"$gte": "/double/bm0146"}}]}]}'

The slk_helpers provide the command gen_file_query with which simple file queries can be generated. The query for the first task could be generated as follows:

slk_helpers gen_file_query '/ex/am/ple/data/.nc$'

Run slk in the “interactive” partition#

If you want to archive/retrieve large files or many files interactively, please use slk archive/retrieve via the interactive partition and allocate 6 GB of memory. This is done by running this command on one of the login nodes of Levante:

salloc --mem=6GB --partition=interactive --account=YOUR_PROJECT_ACCOUNT

Your user account has to be associated to a project account (YOUR_PROJECT_ACCOUT) which as compute time available. Please contact support@dkrz.de if you are not member in such a project but need to use salloc or sbatch for data retrieval or archival.

Run slk as batch job#

A “batch job” denotes a script or a program which is running on one of the Levante compute or shared nodes. The script/program was submitted via sbatch to the SLURM resource manager.

sbatch MY_JOB_SCRIPT.sh

Please always run slk jobs with 6 GB of memory allocated: either via a parameter to the sbatch command (--mem=6GB) or in the header of the batch script (#SBATCH --mem=6GB).

You can check the status of your jobs (queued, running, finishing, …) via

squeue -u $USER

Details on the usage of sbatch and similar SLURM-related commands are given in SLURM Introduction. Exemplary batch scripts for archival and retrieval a given in the sections Archival script templates and Retrieval script templates, respectively.

Your user account has to be associated to a project account (YOUR_PROJECT_ACCOUT) which as compute time available. Please contact support@dkrz.de if you are not member in such a project but need to use salloc or sbatch for data retrieval or archival.

Debugging#

Please have a look into the slk log file in ~/.slk/slk-cli.log for detailed error messages. If you send error reports or questions on failed slk calls to the DKRZ support, please attach your slk log.

pyslk#

Most slk and slk_helpers commands are available as functions pyslk.pyslk.slk_COMMAND(...). These functions are simple wrappers that print out the text, which the slk/slk_helpers commands normally print to the command line. A bit more advanced wrappers are available for a few commands via pyslk.parsers.slk_COMMAND_.... pyslk is installed on Levante in the latest python3 module. It can also be downloaded separately – however, needs slk to be installed. Details on the availability a listed here: pyslk availability.

A few usage examples:

> import pyslk
> pyslk.list('/arch')
    permissions    owner   group  size  day  month  year  filename
0   drwxrwxr-x-  a270003  aa0049         06   Aug   2021  10:17    aa0049
1   drwxrwxr-x-  a270003  aa0238         06   Aug   2021  07:04    aa0238
2   drwxrwx-w--  a270003  ab0036         06   Aug   2021  10:13    ab0036
3   drwxrwx-w--  a270003  ab0051         06   Aug   2021  11:56    ab0051
4   ...
> pyslk.version()
SCLI Version 3.3.21
On this page