Overview
This document describes SLURM node selection plugins and the API that defines
them. It is intended as a resource to programmers wishing to write their own SLURM
node selection plugins. This is version 0 of the API.
SLURM node selection plugins are SLURM plugins that implement the SLURM node selection
API described herein. They are intended to provide a mechanism for both selecting
nodes for pending jobs and performing any system-specific tasks for job launch or
termination. The plugins must conform to the SLURM Plugin API with the following
specifications:
const char plugin_type[]
The major type must be "select." The minor type can be any recognizable
abbreviation for the type of node selection algorithm. We recommend, for example:
- linearA plugin that selects nodes assuming a one-dimensional
array of nodes. The nodes are selected so as to minimize the number of consecutive
sets of nodes utilizing a best-fit algorithm.
- bluegeneIBM Blue Gene
node selector. Note that this plugin not only selects the nodes for a job, but performs
some initialization and termination functions for the job.
The plugin_name and
plugin_version
symbols required by the SLURM Plugin API require no specialization for node selection support.
Note carefully, however, the versioning discussion below.
A simplified flow of logic follows:
slurmctld daemon starts
if (select_p_state_restor)() != SLURM_SUCCESS)
abort
slurmctld reads the rest of its configuration and state information
if (select_p_node_init() != SLURM_SUCCESS)
abort
if (select_p_part_init() != SLURM_SUCCESS)
abort
wait for job
if (select_p_job_test(all available nodes) != SLURM_SUCCESS) {
if (select_p_job_test(all configured nodes) != SLURM_SUCCESS)
reject the job and tell the user it can never run
else
leave the job queued for later execution
} else {
update job's node list and node bitmap
if (select_p_job_init() != SLURM_SUCCESS)
leave the job queued for later execution
else {
execute the job
wait for job to end or be terminated
select_p_job_fini()
}
}
wait for slurmctld shutdown request
select_p_state_save()
Depending upon failure modes, it is possible that
select_p_state_save()
will not be called at slurmctld terminatation.
When slurmctld is restarted, other function calls may be replayed.
select_p_node_init() may be used
to syncronize the plugin's state with that of slurmctld.
Data Objects
These functions are expected to read and/or modify data structures directly in
the slurmctld daemon's memory. Slurmctld is a multi-threaded program with independent
read and write locks on each data structure type. Thererfore the type of operations
permitted on various data structures is identified for each function.
These functions make use of bitmaps corresponding to the nodes in a table.
The function select_p_node_init() should
be used to establish the initial mapping of bitmap entries to nodes.
Functions defined in src/common/bitmap.h should be used for bitmap
manipulations (these functions are directly accessible from the plugin).
API Functions
The following functions must appear. Functions which are not implemented should
be stubbed.
Global Node Selection Functions
int select_p_state_save (char *dir_name);
Description: Save any global node selection state
information to a file within the specified directory. The actual file name used is plugin specific.
It is recommended that the global switch state contain a magic number for validation purposes.
This function is called by the slurmctld deamon on shutdown.
Arguments: dir_name
(input) fully-qualified pathname of a directory into which user SlurmUser (as defined
in slurm.conf) can create a file and write state information into that file. Cannot be NULL.
Returns: SLURM_SUCCESS if successful. On failure,
the plugin should return SLURM_ERROR.
int select_p_state_restore (char *dir_name);
Description: Restore any global node selection state
information from a file within the specified directory. The actual file name used is plugin specific.
It is recommended that any magic number associated with the global switch state be verified.
This function is called by the slurmctld deamon on startup.
Arguments: dir_name
(input) fully-qualified pathname of a directory containing a state information file
from which user SlurmUser (as defined in slurm.conf) can read. Cannot be NULL.
Returns: SLURM_SUCCESS if successful. On failure,
the plugin should return SLURM_ERROR, causing slurmctld to exit.
int select_p_node_init (struct node_record *node_ptr, int node_cnt);
Description: Note the initialization of the node record data
structure. This function is called when the node records are initially established and again
when any nodes are added to or removed from the data structure.
Arguments:
node_ptr (input) pointer
to the node data records. Data in these records can read. Nodes deleted after initiialization
may have their the name field in the record cleared (zero length) rather than
rebuilding the node records and bitmaps.
node_cnt (input) number
of node data records.
Returns: SLURM_SUCCESS if successful. On failure,
the plugin should return SLURM_ERROR, causing slurmctld to exit.
int select_p_part_init (List part_list);
Description: Note the initialization of the partition record data
structure. This function is called when the partition records are initially established and again
when any partition configurations change.
Arguments:
part_list (input) list of partition
record entries. Note that some of these partitions may have no associated nodes. Also
consider that nodes can be removed from one partition and added to a different partition.
Returns: SLURM_SUCCESS if successful. On failure,
the plugin should return SLURM_ERROR, causing slurmctld to exit.
Job-Specific Node Selection Functions
int select_p_job_test (struct job_record *job_ptr,
bitstr_t *bitmap, int min_nodes, int max_nodes);
Description: Given a job's scheduling requirement
specification and a set of nodes which might be used to satisfy the request, identify
the nodes which "best" satify the request. Note that nodes being considered for allocation
to the job may include nodes already allocated to other jobs, even if node sharing is
not permitted. This is done to ascertain whether or not job may be allocated resources
at some later time (when the other jobs complete). This permits SLURM to reject
non-runnable jobs at submit time rather than after they have spent hours queued.
Informing users of problems at job submission time permits them to quickly resubmit
the job with appropriate constraints.
Arguments:
job_ptr (input) pointer
to the job being considered for scheduling. Data in this job record may safely be read.
Data of particular interst include details->contiguous (set if allocated nodes
should be contiguous), num_procs (minimum processors in allocation) and
details->req_node_bitmap (specific required nodes).
bitmap (input/output)
bits representing nodes which might be allocated to the job are set on input.
This function should clear the bits representing nodes not required to satisfy
job's scheduling request.
Bits left set will represent nodes to be used for this job. Note that the job's
required nodes (details->req_node_bitmap) will be a superset
bitmap when the function is called.
min_nodes (input)
minimum number of nodes to allocate to this job. Note this reflects both job
and partition specifications.
max_nodes (input)
maximum number of nodes to allocate to this job. Note this reflects both job
and partition specifications.
Returns: SLURM_SUCCESS if successful. On failure,
the plugin should return SLURM_ERROR and future attempts may be made to schedule
the job.
int select_p_job_init (struct job_record *job_ptr);
Description: Note the initiation of the specified job
is about to begin. This function is called immediately after
select_p_job_test() sucessfully completes for this job.
Arguments:
job_ptr (input) pointer
to the job being initialized. Data in this job record may safely be read or written.
The nodes and node_bitmap fields of this job record identify the
nodes which have already been selected for this job to use. For an example of
a job record field that the plugin may write into, see select_id.
Returns: SLURM_SUCCESS if successful. On failure,
the plugin should return SLURM_ERROR, which causes the job to be requeued for
later execution.
int select_p_job_fini (struct job_record *job_ptr);
Description: Note the termination of the
specified job. This function is called as the termination process for the
job begins (prior to killing the tasks).
Arguments:
job_ptr (input) pointer
to the job being terminated. Data in this job record may safely be read or written.
The nodes and/or node_bitmap fields of this job record identify the
nodes which were selected for this job to use.
Returns: SLURM_SUCCESS if successful. On failure,
the plugin should return SLURM_ERROR.
Versioning
This document describes version 0 of the SLURM node selection API. Future
releases of SLURM may revise this API. A node selection plugin conveys its ability
to implement a particular API version using the mechanism outlined for SLURM plugins.
In addition, the credential is transmitted along with the version number of the
plugin that transmitted it. It is at the discretion of the plugin author whether
to maintain data format compatibility across different versions of the plugin.
|