From c1795de88dba072fda6e75dcd893a363ca8ef0b1 Mon Sep 17 00:00:00 2001
From: Moe Jette <jette1@llnl.gov>
Date: Fri, 31 May 2002 21:37:36 +0000
Subject: [PATCH] Added table of message traffic by client/server. Added "Claim
allocation" message. Made minor changes to others.
---
doc/txt/message.summary.txt | 135 ++++++++++++++++++++++++------------
1 file changed, 90 insertions(+), 45 deletions(-)
diff --git a/doc/txt/message.summary.txt b/doc/txt/message.summary.txt
index f1f7f271a49..04a2598d12f 100644
--- a/doc/txt/message.summary.txt
+++ b/doc/txt/message.summary.txt
@@ -6,8 +6,8 @@ form a REQ/REPLY pair, where - following Moe's convention - the Input is
contained in the `request' message, and the Output will be found in the
`reply.'
-Command(s): Get job information, separate commands for
- accounting, node, partition, job step and build info
+Command(s): Get job/accounting/node/partition/job_step/build information,
+ separate API for each data type
Client: squeue and scontrol commands, plus DPCS from API, any node in cluster
Server: slurmctld
Input: time-stamp, version, user id
@@ -16,94 +16,106 @@ Output: error code, version, time-stamp, record count, array of records
Notes: most information generally available, some might be restricted by user id
-Command(s): Get key
+Command(s): Get partition_key
Client: API call (used by DPCS)
Server: slurmctld
Input: uid (must be root)
-Output: key
+Output: partition_key
Notes: used to control access to some partitions. for example, any user
can run jobs in the "batch" partition, but only when initiated by
a batch controller (e.g. DPCS). this prevents users from running
jobs outside of the queue structure
-Command(s): Allocate
+Command(s): Allocate job
Client: srun or slurm api call
Server: slurmctld
Input: username/uid,nnodes,ntasks, group
- optional: partition,time_limit,constraints,features,node list, key
+ optional: partition,time_limit,constraints,features,node list, partition_key
flags : wait_for_resources, test only (don't allocate resources,
just reply whether or not allocate would have succeeded,
used by DPCS)
-Output: jobid, return code, error code, node list, ncpus for *each* node in list
+Output: job_id, return code, error code, node list, ncpus for *each* node in list,
+ job_key
Notes: allocate resources to a ``job''
-Command(s): Submit
+Command(s): Claim job allocation
+Client: srun
+Server: slurmctld
+Input: uid, job_id, job_key
+Output: error_code
+Notes: ties allocation to a specific process_id, used to determine when a
+ job is really complete
+
+
+Command(s): Submit job
Client: srun or slurm api call
Server: slurmctld
Input: Allocate input + script path, environment, cwd
optional: partition, time_limit, constraints, features,
- I/O location, signal handling, key
+ I/O location, signal handling, partition_key
flags:
-Output: jobid, return code, error code
+Output: job_id, return code, error code
Notes: submit a batch job to the slurm queue
Command(s): Run Job Step
Client: srun or slurm api call
Server: slurmctld
-Input: jobid,username/uid
+Input: job_id,username/uid
optional: nnodes,ntasks,cpus_per_task,distribution,time_limit,
constraints,features,signal handling
flags : wait_for_resources
-Output: jobid, stepid, return code, error code, node list, ncpus/node
+Output: job_id, step_id, return code, error code, node list, ncpus/node
credential list,
Notes: run a set of parallel tasks under an allocated job
- allocate resources if jobid < MIN_JOBID, otherwise assume
+ allocate resources if job_id < MIN_JOBID, otherwise assume
resources are already available
-Command(s): Job Resource Request
+Command(s): Job Step Resource Request
Client: srun, scancel
Server: slurmctld
-Input: stepid
+Input: job_id, step_id, uid
Output: return code, error code, node list, ncpus/node, credentials
Notes: obtain a new set of credentials for a job. Needed for
- at least `srun --attach`
+ at least `srun --attach`, uid must match that of job_id
-Command(s): Run Job Request
+Command(s): Run Job Step Request
Client: srun or slurmctld
Server: slurmd
-Input: username/uid, jobid, stepid, credential, ntasks, environment,
+Input: username/uid, job_id, step_id, credential, ntasks, environment,
cwd, command line, stdin location, stdout/err location
Output: return code, error code
Notes: request initiation of ntasks tasks on this node.
-Command(s): Signal Job Request
+Command(s): Signal Job Step Request
Client: srun or slurmctld (possibly scancel)
Server: slurmd
-Input: uid, jobid or stepid, signal no.
- optional: task no.
+Input: uid, signal no., job_id
+ optional: step_id, task no.
Output: return code
-Notes:
+Notes: Signal all steps and all tasks unless otherwise specified.
+ This could be used to support gang scheduling
-Command(s): Kill Job Request
-Client: srun or slurmctld (possibly scancel)
+Command(s): Kill Job Step Request
+Client: srun or slurmctld or scancel
Server: slurmd
-Input: uid, jobid or stepid
+Input: uid, job_id
+ optional: step_id
Output: return code
-Notes: explicitly kill job as opposed to implicit job kill
- with a signal job request.
+Notes: explicitly kill job as opposed to implicit job kill with a
+ signal job request. Kill all steps unless otherwise specified.
-Command(s): Job Attach Request
+Command(s): Job Step Attach Request
Client: srun
Server: slurmd
-Input: uid, stepid
+Input: uid, job_id, step_id
Output: return code, error code,
stdout/err duplicated to srun stdout/err, signals propagated,
Notes: srun process ``attaches'' to a currently running job. This
@@ -111,12 +123,14 @@ Notes: srun process ``attaches'' to a currently running job. This
to interactively reattach to a batch job.
-Command(s): Cancel job or allocation
+Command(s): Cancel job step or entire job
Client: scancel user command, plus DPCS from API, any node in cluster
Server: slurmctld
-Input: user id, jobid or stepid
+Input: user id, job_id
+ optional: step_id
Output: error code
-Notes: can only be run as user root or the user id for the job
+Notes: Can only be run as user root or the user id for the job
+ Cancel all steps unless otherwise specified.
Command(s): Reconfigure (tell slurmctld to re-read configuration)
@@ -127,15 +141,15 @@ Output: error code, version
Notes: can only be run as user root
-Command(s): Register node (slurmd starting)
+Command(s): Register node
Client: slurmd daemon, any node in cluster
Server: slurmctld
Input: version, time stamp, processor count, memory size, temporary disk space
Output: none
-Notes:
+Notes: Done when slurmd restarts
-Command(s): Status node
+Command(s): Report node status
Client: slurmctld or backup slurmctld
Server: slurmd or slurmctld (for backup check) daemon, any node in cluster
Input: none
@@ -157,16 +171,9 @@ Client: DPCS API
Server: slurmd daemon on the same node as DPCS API is executed
Input: process id
Output: SLURM job id
-Notes: until SLURM accounting is fully funcational, DPCS needs help figuring
- out what processes are associated with each job
-
-
-Command(s): Create job step
-Client: srun from user script, any node in cluster
-Server: slurmctld
-Input: job id, uid, node list, task distribution, processors per task
-Output: step id, elan context (opaque data structure)
-Notes: needed to start parallel program
+Notes: Until SLURM accounting is fully funcational, DPCS needs help figuring
+ out what processes are associated with each job. All message traffic
+ within a node
Command(s): Get job step infomration
@@ -211,6 +218,44 @@ Notes: On termination of a job (not the job step), slurmctld tells
slurmd to execute its epilog program (if any).
+Summary of interactions:
+dpcs->slurmd Get job id from process id
+
+scancel->slurmctld Cancel job step or entire job
+
+scancel->slurmd Kill Job Step Request
+ Signal Job Step Request
+
+scontrol->slurmctld Reconfigure
+ Modify job information
+ Modify node information
+ Modify partition information
+ Get job/accounting/node/partition/job_step/build information
+
+slurmctld->slurmctld Report node status (backup to primary controller)
+
+slurmctld->slurmd Kill Job Step Request
+ Report node status
+ Run epilog
+ Run Job Step Request
+ Upload accounting information
+ Signal Job Step Request
+
+slurmd->slurmctld Get job step infomration
+ Register node
+
+srun->slurmctld Get job step infomration
+ Job Step Attach Request
+ Job Step Resource Request
+ Run Job Step
+ Submit job
+ Allocate job
+ Claim job allocation
+
+srun->slurmd Kill Job Step Request
+ Signal Job Step Request
+ Run Job Step Request
+
----TEMPLATE----
Command(s):
Client:
--
GitLab