_batching

Submodule to determine the number of total/parallel tasks per SLURM job.

_estimate_parallel_tasks_per_job(*, cpus_per_task, mem_per_task, max_cpus_per_job, max_mem_per_job)

Compute how many parallel tasks can fit in a given SLURM job

Note: If more resources than available are requested, return 1. This assumes that further checks will be performed on the output of the current function, as is the case in the heuristics function below.

Parameters:

Name	Type	Description	Default
cpus_per_task	int	Number of CPUs needed for one task.	required
mem_per_task	int	Memory (in MB) needed for one task.	required
max_cpus_per_job	int	Maximum number of CPUs available for one job.	required
max_mem_per_job	int	Maximum memory (in MB) available for one job.	required

Returns:

Type	Description
int	Number of parallel tasks per job

Source code in fractal_server/app/runner/executors/slurm/_batching.py

def _estimate_parallel_tasks_per_job(
    *,
    cpus_per_task: int,
    mem_per_task: int,
    max_cpus_per_job: int,
    max_mem_per_job: int,
) -> int:
    """
    Compute how many parallel tasks can fit in a given SLURM job

    Note: If more resources than available are requested, return 1. This
    assumes that further checks will be performed on the output of the current
    function, as is the case in the `heuristics` function below.

    Arguments:
        cpus_per_task: Number of CPUs needed for one task.
        mem_per_task: Memory (in MB) needed for one task.
        max_cpus_per_job: Maximum number of CPUs available for one job.
        max_mem_per_job: Maximum memory (in MB) available for one job.

    Returns:
        Number of parallel tasks per job
    """
    if cpus_per_task > max_cpus_per_job or mem_per_task > max_mem_per_job:
        return 1
    val_based_on_cpus = max_cpus_per_job // cpus_per_task
    val_based_on_mem = max_mem_per_job // mem_per_task
    return min(val_based_on_cpus, val_based_on_mem)

heuristics(*, tot_tasks, tasks_per_job=None, parallel_tasks_per_job=None, cpus_per_task, mem_per_task, target_cpus_per_job, max_cpus_per_job, target_mem_per_job, max_mem_per_job, target_num_jobs, max_num_jobs)

Heuristically determine parameters for multi-task batching

"In-job queues" refer to the case where parallel_tasks_per_job<tasks_per_job, that is, where not all tasks of a given SLURM job will be executed at the same time.

This function goes through the following branches:

1. Validate/fix parameters, if they are provided as input.
2. Heuristically determine parameters based on the per-task resource requirements and on the target amount of per-job resources, without resorting to in-job queues.
3. Heuristically determine parameters based on the per-task resource requirements and on the maximum amount of per-job resources, without resorting to in-job queues.
4. Heuristically determine parameters (based on the per-task resource requirements and on the maximum amount of per-job resources) and then introduce in-job queues to satisfy the hard constraint on the maximum number of jobs.

Parameters:

Name	Type	Description	Default
tot_tasks	int	Total number of elements to be processed (e.g. number of images in a OME-NGFF array).	required
tasks_per_job	Optional[int]	If tasks_per_job and parallel_tasks_per_job are not None, validate/edit this choice.	None
parallel_tasks_per_job	Optional[int]	If tasks_per_job and parallel_tasks_per_job are not None, validate/edit this choice.	None
cpus_per_task	int	Number of CPUs needed for each parallel task.	required
mem_per_task	int	Memory (in MB) needed for each parallel task.	required
target_cpus_per_job	int	Optimal number of CPUs for each SLURM job.	required
max_cpus_per_job	int	Maximum number of CPUs for each SLURM job.	required
target_mem_per_job	int	Optimal amount of memory (in MB) for each SLURM job.	required
max_mem_per_job	int	Maximum amount of memory (in MB) for each SLURM job.	required
target_num_jobs	int	Optimal total number of SLURM jobs for a given WorkflowTask.	required
max_num_jobs	int	Maximum total number of SLURM jobs for a given WorkflowTask.	required

Return: Valid values of tasks_per_job and parallel_tasks_per_job.

Source code in fractal_server/app/runner/executors/slurm/_batching.py

def heuristics(
    *,
    # Number of parallel components (always known)
    tot_tasks: int,
    # Optional WorkflowTask attributes:
    tasks_per_job: Optional[int] = None,
    parallel_tasks_per_job: Optional[int] = None,
    # Task requirements (multiple possible sources):
    cpus_per_task: int,
    mem_per_task: int,
    # Fractal configuration variables (soft/hard limits):
    target_cpus_per_job: int,
    max_cpus_per_job: int,
    target_mem_per_job: int,  # in MB
    max_mem_per_job: int,  # in MB
    target_num_jobs: int,
    max_num_jobs: int,
) -> tuple[int, int]:
    """
    Heuristically determine parameters for multi-task batching

    "In-job queues" refer to the case where
    `parallel_tasks_per_job<tasks_per_job`, that is, where not all
    tasks of a given SLURM job will be executed at the same time.

    This function goes through the following branches:

    1. Validate/fix parameters, if they are provided as input.
    2. Heuristically determine parameters based on the per-task resource
       requirements and on the target amount of per-job resources, without
       resorting to in-job queues.
    3. Heuristically determine parameters based on the per-task resource
       requirements and on the maximum amount of per-job resources, without
       resorting to in-job queues.
    4. Heuristically determine parameters (based on the per-task resource
       requirements and on the maximum amount of per-job resources) and then
       introduce in-job queues to satisfy the hard constraint on the maximum
       number of jobs.

    Arguments:
        tot_tasks:
            Total number of elements to be processed (e.g. number of images in
            a OME-NGFF array).
        tasks_per_job:
            If `tasks_per_job` and `parallel_tasks_per_job` are not
            `None`, validate/edit this choice.
        parallel_tasks_per_job:
            If `tasks_per_job` and `parallel_tasks_per_job` are not
            `None`, validate/edit this choice.
        cpus_per_task:
            Number of CPUs needed for each parallel task.
        mem_per_task:
            Memory (in MB) needed for each parallel task.
        target_cpus_per_job:
            Optimal number of CPUs for each SLURM job.
        max_cpus_per_job:
            Maximum number of CPUs for each SLURM job.
        target_mem_per_job:
            Optimal amount of memory (in MB) for each SLURM job.
        max_mem_per_job:
            Maximum amount of memory (in MB) for each SLURM job.
        target_num_jobs:
            Optimal total number of SLURM jobs for a given WorkflowTask.
        max_num_jobs:
            Maximum total number of SLURM jobs for a given WorkflowTask.
    Return:
        Valid values of `tasks_per_job` and `parallel_tasks_per_job`.
    """
    # Preliminary checks
    if bool(tasks_per_job) != bool(parallel_tasks_per_job):
        msg = (
            "tasks_per_job and parallel_tasks_per_job must "
            "be both set or both unset"
        )
        logger.error(msg)
        raise SlurmHeuristicsError(msg)
    if cpus_per_task > max_cpus_per_job:
        msg = (
            f"[heuristics] Requested {cpus_per_task=} "
            f"but {max_cpus_per_job=}."
        )
        logger.error(msg)
        raise SlurmHeuristicsError(msg)
    if mem_per_task > max_mem_per_job:
        msg = (
            f"[heuristics] Requested {mem_per_task=} "
            f"but {max_mem_per_job=}."
        )
        logger.error(msg)
        raise SlurmHeuristicsError(msg)

    # Branch 1: validate/update given parameters
    if tasks_per_job and parallel_tasks_per_job:
        # Reduce parallel_tasks_per_job if it exceeds tasks_per_job
        if parallel_tasks_per_job > tasks_per_job:
            logger.warning(
                "[heuristics] Set parallel_tasks_per_job="
                f"tasks_per_job={tasks_per_job}"
            )
            parallel_tasks_per_job = tasks_per_job

        # Check requested cpus_per_job
        cpus_per_job = parallel_tasks_per_job * cpus_per_task
        if cpus_per_job > target_cpus_per_job:
            logger.warning(
                f"[heuristics] Requested {cpus_per_job=} "
                f"but {target_cpus_per_job=}."
            )
        if cpus_per_job > max_cpus_per_job:
            msg = (
                f"[heuristics] Requested {cpus_per_job=} "
                f"but {max_cpus_per_job=}."
            )
            logger.error(msg)
            raise SlurmHeuristicsError(msg)

        # Check requested mem_per_job
        mem_per_job = parallel_tasks_per_job * mem_per_task
        if mem_per_job > target_mem_per_job:
            logger.warning(
                f"[heuristics] Requested {mem_per_job=} "
                f"but {target_mem_per_job=}."
            )
        if mem_per_job > max_mem_per_job:
            msg = (
                f"[heuristics] Requested {mem_per_job=} "
                f"but {max_mem_per_job=}."
            )
            logger.error(msg)
            raise SlurmHeuristicsError(msg)

        # Check number of jobs
        num_jobs = math.ceil(tot_tasks / tasks_per_job)
        if num_jobs > target_num_jobs:
            logger.debug(
                f"[heuristics] Requested {num_jobs=} "
                f"but {target_num_jobs=}."
            )
        if num_jobs > max_num_jobs:
            msg = f"[heuristics] Requested {num_jobs=} but {max_num_jobs=}."
            logger.error(msg)
            raise SlurmHeuristicsError(msg)
        logger.debug("[heuristics] Return from branch 1")
        return (tasks_per_job, parallel_tasks_per_job)

    # 2: Target-resources-based heuristics, without in-job queues
    parallel_tasks_per_job = _estimate_parallel_tasks_per_job(
        cpus_per_task=cpus_per_task,
        mem_per_task=mem_per_task,
        max_cpus_per_job=target_cpus_per_job,
        max_mem_per_job=target_mem_per_job,
    )
    tasks_per_job = parallel_tasks_per_job
    num_jobs = math.ceil(tot_tasks / tasks_per_job)
    if num_jobs <= target_num_jobs:
        logger.debug("[heuristics] Return from branch 2")
        return (tasks_per_job, parallel_tasks_per_job)

    # Branch 3: Max-resources-based heuristics, without in-job queues
    parallel_tasks_per_job = _estimate_parallel_tasks_per_job(
        cpus_per_task=cpus_per_task,
        mem_per_task=mem_per_task,
        max_cpus_per_job=max_cpus_per_job,
        max_mem_per_job=max_mem_per_job,
    )
    tasks_per_job = parallel_tasks_per_job
    num_jobs = math.ceil(tot_tasks / tasks_per_job)
    if num_jobs <= max_num_jobs:
        logger.debug("[heuristics] Return from branch 3")
        return (tasks_per_job, parallel_tasks_per_job)

    # Branch 4: Max-resources-based heuristics, with in-job queues
    parallel_tasks_per_job = _estimate_parallel_tasks_per_job(
        cpus_per_task=cpus_per_task,
        mem_per_task=mem_per_task,
        max_cpus_per_job=max_cpus_per_job,
        max_mem_per_job=max_mem_per_job,
    )
    tasks_per_job = math.ceil(tot_tasks / max_num_jobs)
    logger.debug("[heuristics] Return from branch 4")
    return (tasks_per_job, parallel_tasks_per_job)