block, destroy it correctly.

at least 3.5.0.  This avoids a stack overflow when running jobs on
more than 120k nodes.

Used for generic stack of slurmctld daemon plugins

I run into a problem with slurm-2.5.1 that IDLE nodes can not be
allocated to jobs. This can be reproduced as follows:

First, submit a job with --no-kill option (I have SLURM_EXCLUSIVE set to
allocate nodes exclusively by default). Then set one of the nodes
allocated to the job(cn2) to state DOWN:

srun: error: Node failure on cn2
srun: error: Node failure on cn2
srun: error: cn2: task 0: Killed
^Csrun: interrupt (one more within 1 sec to abort)
srun: task 1: running
srun: task 0: exited abnormally
^Csrun: sending Ctrl-C to job 22605.0
srun: Job step aborted: Waiting up to 2 seconds for job step to finish.
srun: Force Terminated job step 22605.0

Then change state of the node to IDLE again. But it can not be allocated
to jobs:

srun: job 22606 queued and waiting for resources

  JOBID PARTITION     NAME     USER  ST       TIME  NODES
NODELIST(REASON)
  22606      work hostname     root  PD       0:00      1 (Resources)
  22604      work   sbatch     root   R       3:06      1 cn1

NodeName=cn2 Arch=x86_64 CoresPerSocket=8
   CPUAlloc=16 CPUErr=0 CPUTot=16 CPULoad=0.05 Features=abc
   Gres=(null)
   NodeAddr=cn2 NodeHostName=cn2
   OS=Linux RealMemory=30000 Sockets=2 Boards=1
   State=IDLE ThreadsPerCore=1 TmpDisk=0 Weight=1
   BootTime=2012-12-24T15:22:34 SlurmdStartTime=2013-01-14T11:06:32
   CurrentWatts=0 LowestJoules=0 ConsumedJoules=0

I traced and located the problem in select/cons_res. The call sequence
is:

slurmctld/node_mgr.c: update_node() =>
slurmctld/job_mgr.c: kill_running_job_by_node_name() =>
excise_node_from_job() =>
plugins/select/cons_res/select_cons_res.c: select_p_job_resized() =>
_rm_job_from_one_node() => _build_row_bitmaps() =>
common/job_resources: remove_job_from_cores()

If there are other jobs running in the partition, the partition row
bitmap will not be set correctly. In the example above, before
_build_row_bitmaps(), output of _dump_part() is:

[2013-01-19T13:24:56+08:00] part:work rows:1 pri:1
[2013-01-19T13:24:56+08:00]   row0: num_jobs 2: bitmap: 16,32-63

after setting the node down, output of _dump_part() is

[2013-01-19T13:24:56+08:00] part:work rows:1 pri:1
[2013-01-19T13:24:56+08:00]   row0: num_jobs 2: bitmap: 16,32-47

Cores of cn2 are not marked as available. Instead, cores of other nodes
are released. When another job requires the node cn2, the following log
message appears:

[2013-01-19T13:25:03+08:00] debug3: cons_res: _vns: node cn2 busy

I do not understand the design of select/cons_res well and I do not know
how to fix this. But it seems that _build_row_bitmaps() should not be
called, since the job is not removed totally, but only one of the nodes
released.

From Chris Holmes, HP:
After several days of brainstorming and debugging, I have identified
a bug in SLURM 2.5.0rc2, related to the 'tree' topology. It was so
early in the execution of the whole SLURM machinery that it took me
some time to figure it out (say, 100 or 200 jobs showing the issue,
with more or less debugging levels increased and extra
instrumentation, with sometimes an uncertain reliability)...

For every “switch” a bitmap of nodes (seen down by the switch) is
built as the topology is discovered through 'topology.conf'.

There is code in read_config.c, executed when the SLURM control
daemon starts, that reorders the nodes (according to their hostname
by default), while the switches table (ie the bitmaps) has already
being built. To reorder the nodes means that the bitmaps of the switches become wrong.

slurm.MEM_PER_CPU, slurm.NO_VAL, etc.

About a year ago I submitted a modification that you incorporated
into SLURM 2.4, which was to allow an admin to modify a job to use
a QOS even though the user did not have access to the QOS.

However, I must have tested it without having the Accounting set
to enforce QOS's. So, if an admin modifies a job to a QOS they
don't have access to, it will be modified, but the job will result
in a state of InvalidQOS, which is reasonable, since this would
handle the case where a user has their QOS removed. A problem,
however, is that even though the scheduler won't schedule the job,
backfill still will.

One approach would be to fix backfill to be consistent with
the scheduler (which should probably occur regardless), but
my thought would be to modify the scheduler to allow the QOS
as long as it was set by an admin, since that was the intent
of the modification to begin with.

I believe it  would only take a single line to change, just
adding a check on the job_ptr->limit_set_qos, to make sure
it was set by an admin:

                if (job_ptr->qos_id) {
                        slurmdb_association_rec_t *assoc_ptr;
                        assoc_ptr = (slurmdb_association_rec_t *)job_ptr->assoc_ptr;
                        if (assoc_ptr &&
                            !bit_test(assoc_ptr->usage->valid_qos,
                                      job_ptr->qos_id) &&
                            !job_ptr->limit_set_qos) {
                                info("sched: JobId=%u has invalid QOS",
                                        job_ptr->job_id);
                                xfree(job_ptr->state_desc);
                                job_ptr->state_reason = FAIL_QOS;
                                continue;
                        } else if (job_ptr->state_reason == FAIL_QOS) {
                                xfree(job_ptr->state_desc);
                                job_ptr->state_reason = WAIT_NO_REASON;
                        }
                }

Phil

Added "KeepAliveTime" configuration parameter

From Matthieu Hautreux:
TCP_KEEPALIVE addition. No TCP_KEEPALIVE seems to be configured in SLURM TCP
exchanges, thus letting the system potentially deadlocked if a remote host
dissapear and the local host is waiting on a read (the write would result in a
EPIPE or SIGPIPE depending on the masked signals). Adding keepalive with a
relatively large timeout value (5 minutes), could enhance the resilience of
SLURM for unexpected packet/connection loss without too much implication on the
scalability of the solution. The timeout could be configurable in case it is
find too aggresive for particular configurations.

Without this patch, if the first listed partition lacks nodes
with required features the job would be rejected.

While this will validate job at submit time, it results in redundant
looping when scheduling jobs. Working on alternate patch now.

submission.

QoS limits enforcement on the controller side is based on a list of used_limits
per user. When a user is not yet added to the list, which is common when the
controller is restarted and the user has no running jobs, the current logic is
to not check some of the "per user limits" and let the submission succeed.
However, if one of these limits is a zero-valued limit, the check chould
failed as it means that no job should be submitted at all as it would
necessarily result in a crossing of the limit.

This patch ensures that even when a user is not yet present in the per user
used_limits list, the 0-valued limits are correctly treated.

Add PriorityFlags value of "TICKET_BASED".

Correction to CPU allocation count logic in for cores without hyperthreading.

Used to specify the communication protocol to be used for ALPS/BASIL.

allow the use of accounting features like associations, qos and limits but
not keep track of jobs or steps in accounting.

Phase 1 of effort. See "man sbatch" option -a/--array option for details.
Creates job records using sbatch. Reports job arrays using scontrol or
squeue. More work coming soon...

stack.

If sched/backfill starts a job with a QOS having NO_RESERVE and not job
time limit, start it with the partition time limit (or one year if the
partition has no time limit) rather than NO_VAL (140 year time limit);

If a standby job, which in this
case has the NO_RESERVE flag set, is submitted
without a time limit, and is backfilled, it
will get an EndTime waaayyyy into the future.

JobId=99 Name=cmdll
   UserId=eckert(1043) GroupId=eckert(1043)
   Priority=12083 Account=sa QOS=standby
   JobState=RUNNING Reason=None Dependency=(null)
   Requeue=1 Restarts=0 BatchFlag=1 ExitCode=0:0
   RunTime=00:00:14 TimeLimit=12:00:00 TimeMin=N/A
   SubmitTime=2012-12-20T11:49:36 EligibleTime=2012-12-20T11:49:36
   StartTime=2012-12-20T11:49:44 EndTime=2149-01-26T18:16:00

so I looked at the code in /src/plugins/sched/backfill:

                if (job_ptr->start_time <= now) {
                        int rc = _start_job(job_ptr, resv_bitmap);
                        if (qos_ptr && (qos_ptr->flags & QOS_FLAG_NO_RESERVE)){
                                job_ptr->time_limit = orig_time_limit;
                                job_ptr->end_time = job_ptr->start_time +
                                                    (orig_time_limit * 60);

Using the debugger I found that if the job does not have a specified
time limit, the job_ptr->time_limit is equal to NO_VAL when it hits
this code.

2.5.