The parsing of the configuration parameter failed with a prefix
  of "node_features" due to vestigial logic for a plugin type
  of "knl".

Was trying to set NUMA mode as MCNUMA and vise-versa
Also change capmc to specify mode before nids, which seems more robust.

the "fputs" function was aborting, trying to write a NULL string pointer.
Also found the log message was printing the configured and read names
  in the wrong order.

abort() rather than continue if pthread_mutex_ calls fail. better
to die early rather than continue on and risk corruption. mirrors
the (now removed) macro definitions from cbuf/hostlist/list.

If job submit time is right at a second boundary, the test could fail

TEST: 15.37
spawn /home/jette/SLURM/install_smd/bin/salloc --begin now+60 --deadline now+600 --time-min 10 sleep 1^M
salloc: error: Job submit/allocate failed: Requested time limit is invalid (missing or exceeds some limit)^M

FAILURE: batch not submitted with a deadline too short
test15.37 FAILURE

[2016-02-12T15:53:53.005] _valid_job_part: job's min_time greater than deadline (10 > 2016-02-12T16:03:52)
[2016-02-12T15:53:53.005] _slurm_rpc_allocate_resources: Requested time limit is invalid (missing or exceeds some limit)

These are very unlikely to ever occur, but this helps harden the code.

Add logic to read current KNL state information using Intel's syscfg
  command on the KNL node.

Calling _init() on the plugins can have side-effects, and is not safe.

E.g., select/bluegene, switch/nrt, */cray, all make calls to
external APIs that could cause unexpected problems.

This reverts commit 19fc9d94.

Conflicts:
	NEWS

None were real errors, but these changes do harden the code.

Compare the saved clustername to slurmctld's, if there's a mismatch prevent
slurmctld from starting to avoid corruption from separate clusters attempting
to share a single state directory.

Bug 2433.

Include <sys/param.h>; don't use a local MAX_PATH_LEN macro.

3 in node_features work, 1 in new gid cache

Non-GLIBC LIBC's are strict about when these functions are visible.

Rather than keeping the program alive forever due to a bad count
Also eliminate a memory free before its use is done

Conflicts:
	src/slurmctld/read_config.c

chnage from 120 to 60 seconds. If it does not end in time, log it
  and orphan the process. For Cray systems with KNL, process boot
  times are likely to be ~20 minutes.

Eliminate duplicate code (put into function) and only power down node
  when needed.

Bug 2427

Reproduced with:
sacctmgr show user withassoc user= user=<username>

switch/federation superceeded by switch/nrt in a few places.

Remove elan example runqsw.c entirely.

Remove various checks from testsuite. Update documentation.

HAVE_ELAN blocks should also go away.