different file.

Conflicts:
	doc/html/team.shtml

This is a set of tools and utilities to augment Slurm

Conflicts:
	NEWS

job/step completion.

bit_unfmt.

Signed-off-by: Danny Auble <da@schedmd.com>

In the job_submit plugin: Remove all slurmctld locks prior to job_submit()
being called for improved performance. If any slurmctld data structures are
read or modified, add locks directly in the plugin.

Avoid having the hostlist functions initialize memory to zero
since it will be overwritten anyway.

Conflicts:
	NEWS

HealthCheckProgram for nodes in any other states than IDLE.#978

Conflicts:
	NEWS

Unload job tables rather than windows at job end. The table unload
also unloads job tables.

The code in question is allocating a buffer to hold a header and a data payload then forwarding it on a socket.

The first thinko is that it sizes the buffer based on the data payload, which won't be enough for the header+data.  For the second thinko, when it realizes there isn't enough room, it adds the data size to the _capacity_ of the buffer, effectively allocating twice as much memory as needed.  (Aesthetically, I dislike adjusting the 'size' field of the structure to declare the buffer is larger before the buffer is actually made larger, thus the 'new_size' variable.)  This patch corrects those thinkos.

Motivation: As we all discovered with host parsing, xmalloc() of large data can be time-consuming, so avoiding it is preferred.  And in this case avoiding it for 2x the memory needed.  This is also a place where the xrealloc_nz() could be used, since the memory is immediately written into.  As you are probably aware, as peer counts creep into the thousands, this data payload creeps into the megabytes, so this can be a noticeable latency.

An alternate implementation is to guess the max size of the header, and allocate header+data and hope the realloc isn't needed.  I don't do that for a couple reasons.  If the guess is wrong, the worst case is that realloc will end up copying the buffer (full size, yet not without the data payload in it) to a new larger buffer, and I really dislike that prospect.  Better is to assemble the small header, and let realloc copy the small buffer into the huge new space.

Modify the src/common/pack.c functions to use the new memory
management functions and not clear the memory allocated.

Added new internal Slurm functions xmalloc_nz() and xrealloc_nz(), which do
not initialize the allocated memory to zero for improved performance.

super heavy systems to be able to only commit every once in a while instead
of after every transaction.  This is much faster, but can introduce a very
small window in which if something happens that makes the slurmdbd end
irregularly some data loss is possible.

it is much faster than doing a commit every time.

switch/nrt - Do not explicitly unload windows for a job on termination,
only unload its table (which automatically unloads its windows).