Here is the way to reproduce it :
[root@cuzco27 georgioy]# salloc -n64 -N4 --exclusive
salloc: Granted job allocation 8
[root@cuzco27 georgioy]#srun -r 0 -n 30 -N 2 sleep 300&
[root@cuzco27 georgioy]#srun -r 1 -n 40 -N 3 sleep 300&
[root@cuzco27 georgioy]# srun: error: slurm_receive_msg: Zero Bytes were transmitted or received
srun: error: Unable to create job step: Zero Bytes were transmitted or received

batch jobs.

Show this HTML in a new window?
There is a minor problem with the display of partition names in
"sinfo".  Without options, the partition name field displays a
asterisk "*" at the end of the name of the Default partition.  If you
specify a formatting option which contains the %P field specifier with
a width option (e.g., sinfo -o %8P) the asterisk also is appended to
the default partition name.  With no width option, the "%P" displays
the name based on the full length of the name string, however, no "*"
is appended on the default partition name.

The attached patch for version 2.4.0-pre4 corrects the problem so that
the "*" is correctly appended when %P with no width specifier is
used. The patch will also apply to version 2.3.4.

  -Don Albert-

Previously the code would come up with how much memory a PE should have
instead of the memory a node should have.

Fix sched/wiki2 to support job account name, gres, partition name, wckey,
or working directory that contains "#" (a job record separator). Without
this patch, the parsing will probably stop once reaching the "#".

and cons_res or gang scheduling.  The slurm.conf is always used for a node
configuration in this case, and it ignores what the actual hardware is
because in this situation the slurmctld needs to make a bunch of bitmaps
before the slurmd's register.

and time limit where it was previously set by an admin.

slurmdbd accounting and running large amounts of jobs (>50 sec).  Job
information could be corrupted before it had a chance to reach the DBD.

qos.

While safeguards are in place to prevent unauthorized users from extending the
TimeLimit of their running jobs, there were no such restrictions for extending
the EndTime.  This patch adds the same constraints to modifying EndTime that
currently exists for modifying TimeLimit.

Add support for new SchedulerParameters of max_depend_depth defining the
maximum number of jobs to test for circular dependencies (i.e. job A waits
for job B to start and job B waits for job A to start). Default value is
10 jobs.

The problem was conflicting logic in the select/cons_res plugin. Some of the code was trying to get the job the maximum node count in the range while other logic was trying to minimize spreading out of the job across multiple switches. As you note, this problem only happens when a range of node counts is specified and the select/cons_res plugin and the topology/tree plugin and even then it is not easy to reproduce (you included all of the details below).

Quoting Martin.Perry@Bull.com:

> Certain combinations of topology configuration and srun -N option produce
> spurious job rejection with "Requested node configuration is not
> available" with select/cons_res. The following example illustrates the
> problem.
>
> [sulu] (slurm) etc> cat slurm.conf
> ...
> TopologyPlugin=topology/tree
> SelectType=select/cons_res
> SelectTypeParameters=CR_Core
> ...
>
> [sulu] (slurm) etc> cat topology.conf
> SwitchName=s1 Nodes=xna[13-26]
> SwitchName=s2 Nodes=xna[41-45]
> SwitchName=s3 Switches=s[1-2]
>
> [sulu] (slurm) etc> sinfo
> PARTITION AVAIL  TIMELIMIT  NODES  STATE NODELIST
> ...
> jkob         up   infinite      4   idle xna[14,19-20,41]
> ...
>
> [sulu] (slurm) etc> srun -N 2-4 -n 4 -p jkob hostname
> srun: Force Terminated job 79
> srun: error: Unable to allocate resources: Requested node configuration is
> not available
>
> The problem does not occur with select/linear, or topology/none, or if -N
> is omitted, or for certain other values for -N (for example, -N 4-4 and -N
> 2-3 work ok). The problem seems to be in function _eval_nodes_topo in
> src/plugins/select/cons_res/job_test.c. The srun man page states that when
> -N is used, "the job will be allocated as many nodes as possible within
> the range specified and without delaying the initiation of the job."
> Consistent with this description, the requested number of nodes in the
> above example is 4 (req_nodes=4).  However, the code that selects the
> best-fit topology switches appears to make the selection based on the
> minimum required number of nodes (min_nodes=2). It therefore selects
> switch s1.  s1 has only 3 nodes from partition jkob. Since this is fewer
> than req_nodes the job is rejected with the "node configuration" error.
>
> I'm not sure where the code is going wrong.  It could be in the
> calculation of the number of needed nodes in function _enough_nodes.  Or
> it could be in the code that initializes/updates req_nodes or rem_nodes. I
> don't feel confident that I understand the logic well enough to propose a
> fix without introducing a regression.
>
> Regards,
> Martin

When the optional max_time is not specified for --switches=count, the site
max (SchedulerParameters=max_switch_wait=seconds) is used for the job.
Based on patch from Rod Schultz.

Patch by Bill Brophy, Bull.

Patch by Don Lipari, LLNL.
https://github.com/chaos/slurm/commit/4de11bf0a8cd18207a60e7d3e1fa7a6fde0da431

CRAY: Fix support for configuration with SlurmdTimeout=0 (never mark
node that is DOWN in ALPS as DOWN in SLURM).

Improve support for overlapping advanced reservations.
Patch from Bill Brophy, Bull.

task/cgroup: minor job step memcg fixes

The current task/cgroup memory code writes to force_empty at job step
completion and then waits for the release agent to be triggered to
remove the memcg. However, force_empty only causes clean cache pages
to be dropped from the memcg and does not actually move charges to
the parent [1].

This has two unfortunate side-effects. First, pages that can't be
dropped by force_empty are in-use and could stay that way indefinitely
(e.g. system library that is in-use until just after force_empty
completes). Thus, the step memcg never becomes 'empty' and the release
agent is not activated. Second, cached pages that can be freed are
likely associated with the job itself, and those files and libraries
will have to be paged in again for subsequent job steps.

In contrast, calling rmdir(2) on a memcg with no active tasks
causes *all* current charges to move to parent, which is really what
we want in this case. This allows cached libraries and binaries to
stay resident and be associated with the job, and also ensures that
the step memcg is removed immediately as the job step ends.

Thus, this patch replaces the write to force_empty with a call
to xcgroup_delete() on the step memcg, which in turn removes
the memcg with rmdir(2).

The functionality of this patch depends on the previous fix that
uses xcgroup_move_process() to move slurmstepd to the root memcg.
Otherwise, there will be leftover slurmstepd threads in the job
step memcg, and the rmdir will fail with EBUSY.

 [1] Sec 4.3: http://www.kernel.org/doc/Documentation/cgroups/memory.txt

In task_cgroup_memory_fini() the implementation attempts to move
the existing slurmstepd task to the root memory cgroup by writing
the result of getpid(2) to the root memory's 'task' file. This
does not work, however, because slurmstepd is multi-threaded and
thus only the main thread is moved.

This patch replaces the explicit write to 'tasks' with a call to
the new xcgroup_move_process() call, which handles moving all
threads in the process.

This patch adds a helper function to common/xcgroup.c to aid
in moving processes between cgroups. If the cgroups.procs file
is writable then writing the PID to that file is used, as this
method moves all threads in a process atomically.

If cgroups.procs is not writable, then each thread must be moved
individually by walking the /proc/PID/task/ directory and writing
each taskid individually to the 'tasks' file in the cgroup. The
second method is racy if a process is concurrently creating
threads, but it is better than the current method of just moving
one of the process's threads.

It looks like changes were made to the man pages. However, --switch is still used for the info, usage, and help strings. The attached patch fixes those.
Rod Schultz, Bull