Map all nodes in cluster to a single front-end node.
Don't repeat ping/register/kill/etc. RPCs to all pseudo nodes, just the
  front-end.
Treat single message for some RPCs as representing all nodes in the
  cluster: register, ping responce, epilog complete, etc.

request occurs prior to initial node registration.
Node that should have been DRAINED was being set to state DRAINING.

switch (was setting DRAINED nodes in state DRAINING).

option. Use SHOW_ALL as flag.

commands modified. Info filtering added to slurmctld.

* Memory leak in slurm_cred.c, added EVP_MD_CTX_cleanup().
* Pthread stack size too small on AIX. Resulting in stack corruption
  and ugly failure modes. Added slurm_attr_init to macros.h to
  explicitly set the stack size  for all pthreads.
* /dev/urandom not present on AIX, use rand() as needed instead in
  constructing a credential. Used in "srun --join".
* getsockopt(Socket, Level, SO_ERROR, &err, OptionLenght) sometime
  returns an error code of -1. This causes an assert failure in
  slurmd/io.c:_update_error_state().
* Function aliasing is not working on AIX. It is being turned off
  via a variable in config.h and "#if" logic in macros.h and
  slurm_xlator.h.
* dlopen failing if plugins reference any functions not present in
  caller. This may be fixed with the LDFLAG "-Wl,-bgcbypass=1000"
  being added for the slurm commands (avoid garbage collection of
  unused functions).
* read() is sometimes generates EAGAIN error, which was not handled
  in some places.
* vsnprintf() for string NULL is printing "" instead of "(null)"
  as produced by snprintf(). More format printing was added to log.c
  to produce more consistent log messages.
* poll() takes a timeout of -1 for unlimited rather than any negative
  number. Modify logic that was always multiplying by 1000 to convert
  usec to msec.
* getopt_long keyword table was not NULL terminated, resulting in
  segfault with invalid command-line argument in most commands.
* xmalloc module assert failures were not generating a core file.
  Changed "fatal();abort();" to "error();abort();".
* Change msg timeout from 3 sec to 5 sec. Running everything on single
  AIX node was very slow.

than old local functions.

flag the required state saves. Perform the saves using synchronous I/O
from just one thread. Under heavy loads this results in much faster
responsiveness and lowers slurmctld's memory and CPU overhead considerably.

data structure. This eliminates risks associated with re-reading slurm.conf.

state recovery.

  clusters with FastSchedule configured off
* Only return DOWN nodes to service if the reason for them being in
  that state is non-responsiveness and ReturnToService configured on
* Some general code clean-up

fast_schedule == 0 (i.e. the node's value is used directly rather
than the partition's configuration value).

modifications were relatively minor - mostly changes in function
names or arguments.

  (it was inappropriately going to DRAINING state).

Fix update node RPC to handle reason field change without state change.
  State was being handled as type int instead of uint16_t so NO_VAL check
  was not working properly.

sharing via node record of job count (0 | 1) and bitmap of nodes which
permit sharing. Previous logic could permit a job accepting shared nodes
to be scheduled on a node that already had a running job not accepting
shared nodes.

gracefully kill all jobs allocated resources on those nodes,
gracefully kill all pending jobs that require those nodes,
leave pending jobs that exclude those nodes but ignore those nodes.
Added "best_effort" argument to node_name2bitmap() function.
Fix potential memory leak when maui scheduler interface resets
the required nodes.
(gnats:342)

DOWN or DRAIN state and not responding. Changed logic to clear NO_RESPOND
flag before proceeding with logic.

slurmctld restart).

be changed whenever a node responded to a "ping" and other insignificant
events, which resulted in the backfill scheduling running more frequently
than required.

purge the request and job if/when the node changes to state DOWN.

KILL_JOB RPC to the slurmd, repeatedly if needed, but don't confirm it).
(gnats:326)

for node registration, which is relatively rare).

and quitting.

  - changed defs of HAVE_LIBELAN3 to HAVE_ELAN

debug3().

Fatal() no longer calls abort(), but terminates job using exit(1).

node begins to respond again.

This is a problem on PVC, where nodes are regularly DRAINED.

scalability. An arbitrary number of requests may be queued and they
are processed one per second until the queue is empty or pending
requests were last attempted recently (configuration parameters set
to 60 seconds as a minimum retry interval).

and when returned to service went improperly back into state DRAINING
(job counter was inconsistent).

transition to DRAINED state.