Using and Administering

Usage Notes

It is important to know how LoadLeveler keywords and commands behave when you disable the default LoadLeveler scheduling algorithm. LoadLeveler scheduling keywords and commands fall into the following categories:

• Keywords not involved in scheduling decisions are unchanged.
• Keywords kept in the job object or in the machine which are used by the LoadLeveler default scheduler have their values maintained as before and passed to the query API.
• Keywords used only by the LoadLeveler default scheduler have no effect.

The following sections discuss some specific keywords and commands and how they behave when you disable the default LoadLeveler scheduling algorithm.

Job Command File Keywords

class - This value is provided by the query APIs. Machines chosen by ll_start_job must have the class of the job available or the request will be rejected.

dependency - Supported as before. Job objects for which dependency cannot be evaluated (because a previous step has not run) are maintained in the NotQueued state, and attempts to start them via ll_start_job will result in an error. If the dependency is met, ll_start_job can start the proc.

hold - ll_start_job cannot start a job that is in Hold status.

min_processors - ll_start_job must specify at least this number of processors.

max_processors - ll_start_job must specify no more than this number of processors.

preferences - Passed to the query API.

requirements - ll_start_job returns an error if the machine(s) specified do not match the requirements of the job. This includes Disk and Virtual Memory requirements.

startdate - The job remains in the Deferred state until the startdate specified in the job is reached. ll_start_job cannot start a job in the Deferred state.

user_priority - Used in calculating the system priority (as described in How Does a Job's Priority Affect Dispatching Order?). The system priority assigned to the job is available through the query API. No other control of the order in which jobs are run is enforced.

Administration File Keywords

master_node_exclusive is ignored.

master_node_requirement is ignored.

maxidle is supported.

maxjobs is ignored.

maxqueued is supported.

max_jobs_scheduled is ignored.

priority is used to calculate the system priority (where appropriate).

speed is available through the query API.

Configuration File Keywords

MACHPRIO is calculated but is not used.

SYSPRIO is calulated and available to the query API.

MAX_STARTERS is calculated, and if starting the job causes this value to be exceeded, ll_start_job returns an error.

NEGOTIATOR_PARALLEL_DEFER is ignored.

NEGOTIATOR_PARALLEL_HOLD is ignored.

NEGOTIATOR_RESCAN_QUEUE is ignored.

NEGOTIATOR_RECALCULATE_SYSPRIO_INTERVAL works as before. Set this value to 0 if you do not want the SYSPRIOs of job objects recalulated.