|SCHED_CONF(5)||Grid Engine File Formats||SCHED_CONF(5)|
sched_conf - Grid Engine default scheduler configuration file
sched_conf defines the configuration file format for Grid Engine's scheduler. In order to modify the configuration, use the graphical user's interface qmon(1) or the -msconf option of the qconf(1) command. A default configuration is provided with the Grid Engine distribution package.
Note, Grid Engine allows backslashes (\) be used to escape newline characters. The backslash and the newline are replaced with a space character before any interpretation.
The following parameters are recognized by the Grid Engine scheduler if present in sched_conf:
Note: Deprecated, may be removed in future release.
Allows for the selection of alternative scheduling algorithms.
Currently default is the only allowed setting.
A simple algebraic expression used to derive a single weighted load value from all or part of the load parameters reported by sge_execd(8) for each host and from all or part of the consumable resources (see complex(5)) being maintained for each host. The load formula expression syntax is that of a sum of weighted load values, that is:
Note, no blanks are allowed in the load formula.
The load values and consumable resources (load_val1, ...) are specified by the name defined in the complex (see complex(5)).
Note: Administrator-defined load values (see the load_sensor parameter in sge_conf(5) for details) and consumable resources available for all hosts (see complex(5)) may be used as well as Grid Engine default load parameters.
The weighting factors (w1, ...) are positive integers. After the expression is evaluated for each host the results are assigned to the hosts and are used to sort the hosts corresponding to the weighted load. The sorted host list is used to sort queues subsequently.
The default load formula is np_load_avg.
The load which is imposed by the Grid Engine jobs running on a
system varies in time, and often, e.g. for the CPU load, requires some
amount of time to be reported in the appropriate quantity by the operating
system. Consequently, if a job was started very recently, the reported load
may not provide a sufficient representation of the load which is already
imposed on that host by the job. The reported load will adapt to the real
load over time, but the period of time in which the reported load is too low
may already lead to an oversubscription of that host. Grid Engine allows the
administrator to specify job_load_adjustments which are used in the
Grid Engine scheduler to compensate for this problem.
The job_load_adjustments are specified as a comma-separated list of arbitrary load parameters or consumable resources and (separated by an equal sign) an associated load correction value. Whenever a job is dispatched to a host by the scheduler, the load parameter and consumable value set of that host is increased by the values provided in the job_load_adjustments list. These correction values are decayed linearly over time until after load_adjustment_decay_time from the start the corrections reach the value 0. If the job_load_adjustments list is assigned the special denominator NONE, no load corrections are performed.
The adjusted load and consumable values are used to compute the combined and weighted load of the hosts with the load_formula (see above) and to compare the load and consumable values against the load threshold lists defined in the queue configurations (see queue_conf(5)). If the load_formula consists simply of the default CPU load average parameter np_load_avg, and if the jobs are very compute intensive, one might want to set the job_load_adjustments list to np_load_avg=1.00, which means that every new job dispatched to a host will require 100% CPU time, and thus the machine's load is instantly increased by 1.00.
The load corrections in the "job_load_adjustments" list above are decayed linearly over time from the point of the job start, where the corresponding load or consumable parameter is raised by the full correction value, until after a time period of "load_adjustment_decay_time" the correction becomes 0. Proper values for "load_adjustment_decay_time" greatly depend upon the load or consumable parameters used and the specific operating system(s). Therefore, they can only be determined on-site and experimentally. For the default np_load_avg load parameter a "load_adjustment_decay_time" of 7 minutes has proven to yield reasonable results.
The maximum number of jobs any user may have running in a Grid Engine cluster at the same time. If set to 0 (default) the users may run an arbitrary number of jobs.
At the time the scheduler thread initially registers with the
event master thread in the sge_qmaster(8) process
schedule_interval is used to set the time interval in which the event
master thread sends scheduling event updates to the scheduler thread. A
scheduling event is a status change that has occurred within
sge_qmaster(8) which may trigger or affect scheduler decisions (e.g.
a job has finished and thus the allocated resources are available again).
In the Grid Engine default scheduler the arrival of a scheduling event report triggers a scheduler run. The scheduler waits for event reports otherwise.
Schedule_interval is a time value (see sge_types(5) for a definition of the syntax of time values). Setting it to 0 disables scheduling.
This parameter determines in which order several criteria are
taken into account to produce a sorted queue instance list which determines
the preferred order for scheduling tasks to them (typically determining the
order in which hosts are used). Currently, two settings are valid:
seqno and load. However in both cases, Grid Engine attempts to
maximize the number of soft requests (see qsub(1) -s option)
being fulfilled by the queues for a particular job as the primary criterion.
Then, if the queue_sort_method parameter is set to seqno, Grid Engine will use the seq_no parameter as configured in the current queue configurations (see queue_conf(5)) as the next criterion to sort the queue list. The load_formula (see above) is only used as the next criterion if two queues have equal sequence numbers. If queue_sort_method is set to load the load according the load_formula is the criterion after maximizing a job's soft requests, and the sequence number is only used if two hosts have the same load. The sequence number sorting is most useful if you want to define a fixed order in which queues are to be filled (e.g. the cheapest resource first).
The default for this parameter is load.
When executing under a share based policy, the scheduler
"ages" (i.e. decreases) usage to implement a sliding window for
achieving the share entitlements as defined by the share tree. The
halftime defines the time interval in which accumulated usage will
have been decayed to half its value at the start of the interval. (This is a
radioactive-type exponential decay, where the parameter is usually called
"half-life".) Valid values are specified in hours, default 168.
If the value is set to 0, the usage is not decayed.
Grid Engine accounts for the consumption of the resources CPU-time, memory and IO to determine the usage which is imposed on a system by a job. A single usage value is computed from these three input parameters by multiplying the individual values by weights and adding them up. The weights are defined in the usage_weight_list. The format of the list is
where wcpu, wmem and wio are the configurable weights. The weights are real numbers. The sum of all three weights should be 1. The default is cpu=1,mem=0,io=0.
Determines how fast Grid Engine should compensate for past usage below or above the share entitlement defined in the share tree. Recommended values are between 2 and 10, where 10 means faster compensation. The default is 5.
The relative importance of the user shares in the functional policy. Values are of type real.
The relative importance of the project shares in the functional policy. Values are of type real.
The relative importance of the department shares in the functional policy. Values are of type real.
The relative importance of the job shares in the functional policy. Values are of type real.
The maximum number of functional tickets available for distribution by Grid Engine. Determines the relative importance of the functional policy. See under sge_priority(5) for an overview on job priorities.
The maximum number of share based tickets available for distribution by Grid Engine. Determines the relative importance of the share tree policy. See under sge_priority(5) for an overview on job priorities.
The weight applied on the remaining time until a job's latest start time. Determines the relative importance of the deadline. See under sge_priority(5) for an overview on job priorities.
The weight applied on the job's waiting time since submission. Determines the relative importance of the waiting time. See under sge_priority(5) for an overview on job priorities.
The weight applied on jobs' normalized urgency when determining the priority finally used. Determines the relative importance of urgency. See under sge_priority(5) for an overview on job priorities.
The weight applied on jobs' normalized POSIX priority when determining the priority finally used. Determines the relative importance of POSIX priority. See under sge_priority(5) for an overview on job priorities.
The weight applied on the normalized ticket amount when determining the priority finally used. Determines the relative importance of the ticket policies. See under sge_priority(5) for an overview on job priorities.
This parameter is provided for tuning the system's scheduling behavior. By default, a scheduler run is triggered in the scheduler interval. When this parameter is set to 1 or larger, the scheduler will be triggered that number of seconds after a job has finished. Setting this parameter to 0 disables the flush after a job has finished.
This parameter is provided for tuning the system's scheduling behavior. By default, a scheduler run is triggered in the scheduler interval. When this parameter is set to 1 or larger, the scheduler will be triggered that number of seconds after a job was submitted to the system. Setting this parameter to 0 disables the flush after a job was submitted.
The default scheduler can keep track of why jobs could not be scheduled during the last scheduler run. This parameter enables or disables the observation. The value true enables the monitoring false turns it off.
It is also possible to activate the observation only for certain jobs. This will be done if the parameter is set to job_list followed by a comma-separated list of job ids.
The user can obtain the collected information with the command qstat -j.
This is for passing additional parameters to the Grid Engine scheduler. The following values are recognized:
- If set, overrides the default of value 60 seconds. This parameter is used
by the Grid Engine scheduler when planning resource utilization as the
delta between net job runtimes and total time until resources become
available again. Net job runtime as specified with -l h_rt=... or
-l s_rt=... or default_duration always differs from total
job runtime due to delays before and after actual job start and finish.
Among the delays before job start is the time until the end of a
schedule_interval, the time it takes to deliver a job to
sge_execd(8), and the delays caused by prolog in
queue_conf(5), start_proc_args in sge_pe(5) and
starter_method in queue_conf(5). The delays after job finish
include those due to a forced job termination (notify,
terminate_method or checkpointing), procedures run after
actual job finish, such as stop_proc_args in sge_pe(5) or
epilog in queue_conf(5), and the delay until a new
If the offset is too low, resource reservations (see max_reservation) can be delayed repeatedly due to an overly optimistic job circulation time.
- Note: Deprecated, may be removed in future release.
If set to true, the scheduler limits the number of jobs it looks at during a scheduling run. At the beginning of the scheduling run it assigns each job a specific category, which is based on the job's requests, priority settings, and the job owner. All scheduling policies will assign the same importance to each job in one category. Therefore the number of jobs per category has a FIFO order and can be limited to the number of free slots in the system.
An exception is jobs which request a resource reservation. They are included regardless of the number of jobs in a category.
This setting is turned off by default, because in very rare cases, the scheduler can make a wrong decision. It is also advised to turn report_pjob_tickets off. Otherwise qstat -ext can report outdated ticket amounts. The information shown with a qstat -j for a job that was excluded in a scheduling run is very limited.
- If set equal to 1, the scheduler logs profiling information summarizing each scheduling run.
- If set equal to 1, the scheduler records information for each scheduling run, enabling reproduction of job resource utilization in the file <sge_root>/<cell>/common/schedule.
- This parameter sets the algorithm for the PE range computation. The
default is automatic, which means that the scheduler will select the best
one, and it should not be necessary to change it to a different setting in
normal operation. If a custom setting is needed, the following values are
auto: the scheduler selects the best algorithm
least: starts the resource matching with the lowest slot amount first
bin: starts the resource matching in the middle of the pe slot range
highest: starts the resource matching with the highest slot amount first.
Changing params will take immediate effect. The default for params is none.
Interval (HH:MM:SS) to reprioritize jobs on the execution hosts based on the current ticket amount for the running jobs. If the interval is set to 00:00:00 the reprioritization is turned off. The default value is 00:00:00. The reprioritization tickets are calculated by the scheduler and update events for running jobs are only sent after the scheduler calculated new values. How often the scheduler should calculate the tickets is defined by the reprioritize_interval. Because the scheduler is only triggered in a specific interval (scheduler_interval) this means the reprioritize_interval only has a meaning if set greater than the scheduler_interval. For example, if the scheduler_interval is 2 minutes and reprioritize_interval is set to 10 seconds, this means the jobs get re-prioritized every 2 minutes.
This parameter allows tuning the system's scheduling run time. It
is used to enable/disable the reporting of pending job tickets to the
qmaster. It does not influence the tickets calculation. The sort order of
jobs in qstat and qmon is only based on the submit time when the reporting
is turned off.
The reporting should be turned off in a system with a very large amount of jobs by setting this parameter to "false".
The halflife_decay_list allows configuring different decay rates
for the finished_jobs usage types, which is used in the pending job ticket
calculation to account for jobs which have just ended. This allows the user
the pending jobs algorithm to count finished jobs against a user or project
for a configurable decayed time period. This feature is turned off by
default, and the halftime is used instead.
The halflife_decay_list also allows one to configure different decay rates for each usage type being tracked (cpu, io, and mem). The list is specified in the following format:
usage_type can be one of cpu, io, or mem. time can be -1, 0 or a timespan specified in minutes. If time is -1, only the usage of currently running jobs is used. 0 means that the usage is not decayed.
This parameter sets up a dependency chain of ticket-based policies. Each ticket-based policy in the dependency chain is influenced by the previous policies and influences the following policies. A typical scenario is to assign precedence for the override policy over the share-based policy. The override policy determines in such a case how share-based tickets are assigned among jobs of the same user or project. Note that all policies contribute to the ticket amount assigned to a particular job regardless of the policy hierarchy definition. Yet the tickets calculated in each of the policies can be different, depending on "POLICY_HIERARCHY".
The "POLICY_HIERARCHY" parameter can be an up to 3 letter combination of the first letters of the 3 ticket based policies S(hare-based), F(unctional) and O(verride). So a value "OFS" means that the override policy takes precedence over the functional policy, which finally influences the share-based policy. Less than 3 letters means that some of the policies do not influence other policies and also are not influenced by other policies. So a value of "FS" means that the functional policy influences the share-based policy and that there is no interference with the other policies.
The special value "NONE" switches off policy hierarchies.
If set to "true" or "1", override tickets of any override object instance are shared equally among all running jobs associated with the object. The pending jobs will get as many override tickets, as they would have, when they were running. If set to "false" or "0", each job gets the full value of the override tickets associated with the object. The default value is "true".
If set to "true" or "1", functional shares of any functional object instance are shared among all the jobs associated with the object. If set to "false" or "0", each job associated with a functional object, gets the full functional shares of that object. The default value is "true".
The maximum number of pending jobs to schedule in the functional policy. The default value is 200.
The maximum number of subtasks per pending array job to schedule. This parameter exists in order to reduce scheduling overhead. The default value is 50.
The maximum number of reservations scheduled within a schedule interval.
When a runnable job can not be started due to a shortage of resources a reservation can be scheduled instead. A reservation can cover consumable resources with the global host, any execution host, and any queue. For parallel jobs reservations are done also for the slots resource as specified in sge_pe(5). The top max_reservation jobs (in priority order) are considered, not individual resources. The job runtime assumed is the maximum of the time specified with -l h_rt=... or -l s_rt=... For jobs that have neither of them, the default_duration (see below) is assumed.
Reservations prevent jobs of lower priority as specified in sge_priority(5) from utilizing the reserved resource quota during the time of reservation. Jobs of lower priority are allowed to utilize those reserved resources only if their prospective job end is before the start of the reservation ("backfilling"). Reservation is done only for non-immediate jobs (-now no) that request reservation (-R y). If max_reservation is set to "0" no job reservation is done.
max_reservation actually has a more general effect on scheduler look-ahead, and it is necessary to turn it on for correct backfilling into calendar windows (see calendar_conf(5)).
Note that reservation scheduling can be performance consuming and hence reservation scheduling is switched off by default. Since reservation scheduling performance consumption is known to grow with the number of pending jobs, the use of the -R y option is recommended only for those jobs actually queuing for bottleneck resources. Together with the max_reservation parameter, this technique can be used to narrow down performance impacts. A JSV can be used to add reservation requests for particular resources, such as large parallel jobs.
When job reservation is enabled through the max_reservation sched_conf(5) parameter, the default_duration is assumed as runtime for jobs that have neither -l h_rt=... nor -l s_rt=... specified. In contrast to an h_rt/s_rt time limit, the default_duration is not enforced. The default value is INFINITY, and reservation is not effective for jobs which get that value, i.e. the value must be finite, or jobs must specify a run time.
<sge_root>/<cell>/common/sched_configuration scheduler thread configuration
See sge_intro(1) for a full statement of rights and permissions.