DESIGN.md

Row

Row is yet another workflow engine that automates the process of executing actions on groups of directories.

Use-cases and implementation notes

• Operate entirely on the file system with no database backend.
• Define an arbitrary number of actions.
  • Track which actions have been completed on each directory.
  • Mark actions complete when all their product files are present.
  • Indicate the previous actions must be completed prior to executing another action.
• Define resources used by an action.
  • Make this information to the job schedulers and launchers.
• Generate a shell script (a job) that executes the action's command on a *group.
  • Optionally use one or more launchers.
  • Allow the user to pass additional launcher arguments.
• Submit jobs to schedulers (e.g. SLURM).
  • Set typical options from convenience keys (account, walltime, partition).
  • Auto-detect partition based on resources (e.g. gpu vs cpu).
  • Set user-defined submission options.
  • Set user-defined setup commands.
  • Automatically submit jobs to the scheduler (e.g. sbatch).
  • Track which actions have been submitted on each directory.
• "Submit" jobs to the shell.
  • Run in the local process with a progress bar.
• Define groups by:
  • Start from all directories in the workspace.
  • When set, include only directories that match a list of pointer-operation-value expressions.
  • Sort by directory name first to provide a stable starting point.
  • Sort by a user-defined list of pointers (if set).
  • Optionally split the sorted sequence into groups where the sort keys are identical.
  • Further split groups down to a user-defined maximum size (if set).
• Define submission groups by additionally:
  • Start from the set of directories given to the command, or all directories if not provided.
  • After filtering but before sorting and grouping, determine eligible directories filtering out:
    • Directories where this action is complete.
    • Directories where the previous actions are not complete.
    • Directories where this action has been submitted to the scheduler.
  • Optionally require whole groups.
• Print the workflow status. For each action, note the number of directories:
  • Queued in SLURM.
  • Running in SLURM. (OR, just submitted).
  • Eligible to run.
  • Waiting on previous actions.
• List directories and show completion status, submitted job ID, and user-defined keys from the value.

Ideas:

• List scheduler jobs and show useful information.
• Cancel scheduler jobs specific to actions and/or directories.

Overview

Row is written in Rust to provide a fast and responsive interface for users with workflows of all sizes. The tool will often be used on HPC resources where filesystem operations are expensive. This design strives to avoid O(N) filesystem operations whenever possible. The Row library implements the types and functions needed for the business logic. The binary parses arguments and dispatches calls to the library.

Library module structure

• row
  • cli - Top level command line commands.
  • cluster - Read the clusters.toml configuration file, determine the active cluster, and make the settings available.
  • launcher - Read the launchers.toml configuration file, provide code to construct commands with launcher prefixes.
  • project - Combine the workflow, state, and scheduler into one object and provide methods that work with the project as a whole.
  • scheduler - Generic scheduler interface and implementations for shell and SLURM.
  • state - Row's internal state of the workspace.
  • workflow - Type defining workflow and utility functions to find and read it.
  • workspace - Methods for working with workspace directories on the filesystem.

Details

Workflow definition

A directory that includes the file workflow.toml is the root of a Row project. When executed, Row checks in the current working directory (and recursively in parent directories) until workflow.toml is found. workflow.toml defines:

• The workspace
  • path
  • A static value file
• cluster-specific submit_options
  • account
  • options
  • setup script
• The actions:
  • name
  • command
  • launchers
  • previous_actions
  • products
  • resources
    • processes (either per_submission or per_directory)
    • threads_per_process
    • gpus_per_process
    • walltime (either per_submission or per_directory)
  • Cluster- and action-specific submit_options
    • options
    • setup
    • partition
  • group
    • include
    • maximum_size
    • submit_whole
    • sort_by
    • split_by_sort_key

Detecting completion

After executing an action, Row checks for the file that it produces. If that file exists, Row records that in the state. In this way, the completion check is performed in parallel on the compute nodes while the job's files are likely still in the filesystem cache. It also allows for restartable actions to take require many executions before completing.

Completion staging files

To mark an action complete, Row writes the list of completed directories to .row/completed/ <random-filename> in a binary file format. This gathers all the completion information without generating conflicts when multiple jobs attempt to record completion information simultaneously.

Row generates the file in a temporary location and rename it to ensure that the entire set of completed directories is read.

The cache files

Row maintains the state of the workflow in several files:

• directories.json
  • Last time the workspace was modified.
  • Cached copies of the user-provided static value file.
• completed.postcard
  • Completion status for each action.
• submitted.postcard
  • The last submitted job ID, referenced by action, directory, and cluster.

When Row updates the state, it collects the completion staging files and updates the entries in the state accordingly. It also checks the scheduler for all known job IDs and removes any job IDs that are no longer running. After writing the new state to disk, it removes the completion staging files. When Row discovers new directories, it adds them to the state and populates the cached static value files. It also removes any directories that are no longer present.

For consistency, Row performs these updates every time it computes the status, is asked to submit jobs, or any command that queries the status. By design the scan command does not update the caches.

Status report

The status report gives an overview of all defined actions at the directory level. Status is unaware of and does not provide information regarding individual groups or jobs. Each time status is called, it updates the state.

In addition to the numbers of directories, the status report also summarizes the total number of CPU (or GPU) hours that eligible and waiting actions would use.

Submission

A submission examines all actions and directories and determines which groups are eligible to submit to the job scheduler. It then prints a status of the expected total resource usage and asks for confirmation before submitting the job(s) to the scheduler. After submitting, Row updates the state with a record of which scheduler job IDs were submitted for each action/directory combination.

Provide a --dry-run option that shows the user what job script(s) would be submitted.

End the remaining submission sequence on an error return value from the scheduler. Save the cache to capture the state of any successfully submitted jobs.

Shell "submissions"

Row is NOT a scheduler itself. It will submit and run whatever the user asks it to. For shell submissions, Row ignores the resource request and assumes the users knows that what they are submitting is reasonable. For scheduler submissions, Row may make some basic sanity checks (e.g. underpopulated whole node submissions) but will otherwise pass through whatever script as generated by user input.

Group definition

The group defining options include and sort_by use JSON pointer syntax. This allows users to select any element of their value when defining groups.

Launcher configuration

Launchers define prefixes that go in front of commands. These prefixes (e.g. OMP_NUM_THREADS, srun) take arguments when the user requests certain resources. Row provides built in support for OpenMP and MPI on the built-in clusters that **row supports. Users can override these and provide new launchers in launchers.toml.

Each launcher optionally emits an executable, and arguments for

• total number of processes
• threads per process
• gpus per process when both the launcher defines such an argument and the user requests the relevant resource.

Cluster configuration

Row provides default configurations for many national HPC systems. Users can override these defaults or define new systems in $HOME/.config/row/clusters.toml.

A single cluster defines:

• name
• identify: one of
  • by_environment: [string, string]
  • always: bool
• scheduler
• partition (listed in priority order)
  • name
  • maximum_cpus_per_job
  • require_cpus_multiple_of
  • maximum_gpus_per_job
  • require_gpus_multiple_of

The list of partitions provides a mechanism by which the scheduler can determine which partition to execute on. It will check the partitions in order and return the first partition where a given job is under both the maximum CPUs and GPUs. Users can override the partition selection in the action definition.

The require_[c/g]pus_multiple_of keys are optional. When set, submit will ensure that each job submitted to this partition has an exact multiple of the given number of CPUs or GPUs. This ensures that users make full use of whole node partitions. Shared partitions should not set these keys.

On creation, a Cluster object instantiates the appropriate Scheduler and passes it the details of the launchers and the partitions. Users can submit to bash directly with the none cluster.

Environment variables

When executing actions, set environment variables so that the action script is aware of the context in which it is running:

• ACTION_CLUSTER - current cluster name.
• ACTION_NAME - the action's name.
• ACTION_PROCESSES_PER_DIRECTORY - the number of processes requested per directory. Only set when action.resources.processes is per_directory.
• ACTION_PROCESSES - the total number of processes in this submission.
• ACTION_THREADS_PER_PROCESS - the number of threads per process. Not set when action.resources.threads_per_process is missing from the action's definition.
• ACTION_GPUS_PER_PROCESS - the number of GPUs per process. Not set when action.resources.gpus_per_process is missing from the action's definition.
• ACTION_WALLTIME_IN_MINUTES - the total walltime in minutes.

Designing actions for resiliency.

Users should implement idempotent actions. For example, an action that generates a file should first check if that file exists. When it does, return immediately. If a user loses or corrupts the state file, then rerunning parts of the workflow will not corrupt the workspace.

Users can manually recheck action completion whenever needed. For example, if the state file is lost, corrupted, or a job is killed before completing successfully. Users will use the same command that row uses internally to check completion status. In large workspaces, checking the completion status may take a long time, so it should display a progress bar.

Subcommands

• init - create workflow.toml and workspace if they do not yet exist.
• scan - scan the workspace for directories that have completed actions.
• show - show properties of the workflow:
  • status - summarize the status of the workflow.
  • directories - list directories in the workflow.
  • clsuter - show the currently selected cluster configuration.
  • launchers - list the launchers for the current cluster.
• clean - delete row cache files.

Definitions

• action: A user-defined script.
• complete: An action is complete after its command has been executed on the directory.
• command: A shell command.
• condition: An expression that determines whether a given directory is a member of a
• directory: A directory on the file system inside the workspace.
• eligible: An action is eligible to execute on a directory when all previous actions are complete, the action is not yet submitted to the scheduler, and the action itself is not yet complete.
• group: A ordered collection of workspace directories.
• job: A shell script that executes an action on a *group.
• launcher: A tool that facilitates executing commands (e.g. srun).
• value file: A JSON, file that stores the value of this directory.
• previous actions: The actions that must all be completed prior to starting a given action.
• product - A file present in the directory after the action is complete.
• project: A filesystem directory that contains a workflow definition and a workspace.
• resources: The number of processes, cpus per process, gpus per process, and walltime needed.
• status: The overall state of the workflow.
• submission group: A group that is ready for a given action to be applied.
• whole group: A submission group that is identical to the group found without applying the additional submission filters.
• workspace: The location on the file system that contains directories.

TODO: logo