On Cray, wait_job means to confirm the already existing ALPS reservation. This
is handled already:
 * for salloc by select_g_job_ready() - hence no need to call again;
 * for batch jobs it is done in the stepdmanager.
Hence just print a warning to the user.

13_scontrol-no-wait_job.diff

This adds support for execution of salloc on a local Cray system,
disabling node sharing (still not supported on XT/XE).

It further disables running salloc within salloc, as it leads to errors: since
Cray uses process group / PAGG IDs for tracking its reservations, running
salloc from within salloc invariably leads to a ALPS resource allocation error.

Thirdly, it disable Cray node allocation on non-Cray systems, since this
requires that the host on which salloc spawns the shell process is capable
of Cray task launch.

If it is not, then the remote slurmctld will reserve the requested nodes, but
the local host runninc salloc will neither be able to confirm the ALPS 
reservation (due to the absence of a local apbasil command), nor would it be
able to run jobs on the compute nodes.

To distinguish this case from general task launch (we use a frontend host where
salloc could end up running jobs on different clusters, depending on the value
exported via $SLURM_CONF), the following condition is tested:

 * Cray build support has been enabled (HAVE_CRAY);
 * the loaded slurm.conf uses select/cray (required on Cray hosts);
 * the local host does not have support for apbasil (HAVE_NATIVE_CRAY undefined).

Since the 'apbasil' command is only available on native Cray systems, this
combination of conditions seems sufficient to prevent accidentally using
salloc on a host which does not support it.

(For sbatch the case is different, since the job script runs on the remote host.)

11_salloc.diff
done with minor change for Cray emulation

This puts the Basil inventory immediately before each (backfill) schedule. 

Having considered multiple alternatives, this is the most robust and least
wasteful solution. The reason is that ALPS keeps internal node state, which
can be changed
 * by the administrator (xtprocadmin),
 * by the node health checker programs (setting some nodes into 'suspect'),
 * by ALPS itself.

Tracking this periodically, e.g. every HealthCheckInterval, may mean to miss
some state changes. The result would not be a crash, but a subsequently
failed ALPS reservation, which would require to undo some of the slurm state.

Also added inventory to plugin/sched/wiki and wiki2 at get_node time

09_Cray-INVENTORY-directly-before-schedule.diff

select/cray: update compile-time and runtime support for Cray build

These changes update build support for Cray XT/XE:
 1. renamed '--cray-xt' into '--cray' since also XE systems are supported;
 2. autoconf rules to cover the various possible build cases:
    a) --enable-cray=off: HAVE_CRAY/HAVE_NATIVE_CRAY undefined,
    b) --enable-cray=on:  HAVE_CRAY defined
       b1) local host is a native Cray system: HAVE_NATIVE_CRAY defined
           (requires installation of mysql-devel and libexpat-devel packages),
       b2) local host is not a native Cray system: the conditionally built
           parts (basil_interface.c, libalps.la) are not built;
 3. updated configure logic:
    - since Cray support depends on mySQL, reordered tests in configure.ac,
    - reordered logic with regard to changes in (2),
    - an AM_CONDITIONAL to build native-Cray parts conditionally,
    - updated configure messages (XT/XE);
 4. run-time read_conf test to ensure use of select/cray is properly supported,
 5. an update of the NEWS file due to the change in (1) ==> may have a conflict
    in case you have a locally-updated copy.

I have compile-tested the three possible scenarios in (2).

    03_Bug-fix_slurmctld-swap-both-NodeAddr-and-NodeHostname-when-reordering.diff

scontrol: warn user that base node state can not be changed on Cray

The base node state (UP, DOWN, ALLOCATED, ...) is handled by ALPS and inferred
from reading the output of ALPS inventory requests.

To avoid inconsistencies, it is not possible for a user to alter this node state.
This patch adds a warning to scontrol if a user wants to change node state through
slurm:

palu> scontrol update NodeName=nid00171 State=DOWN
State=DOWN can not be changed through slurm: use native Cray tools such as e.g. xtprocadmin(8)

The 'meta' states such as DRAIN can still be changed.

This reverses some patches from Gerrit that were old, going to work
forward now from the start

    XT/XE systems, and build on native Cray XT/XE systems (auto-detected).
    Building on native Cray systems requires the cray-MySQL-devel-enterprise
    rpm and expat XML parser library/headers.

select/cray: update compile-time and runtime support for Cray build

These changes update build support for Cray XT/XE:
 1. renamed '--cray-xt' into '--cray' since also XE systems are supported;
 2. autoconf rules to cover the various possible build cases:
    a) --enable-cray=off: HAVE_CRAY/HAVE_NATIVE_CRAY undefined,
    b) --enable-cray=on:  HAVE_CRAY defined
       b1) local host is a native Cray system: HAVE_NATIVE_CRAY defined
           (requires installation of mysql-devel and libexpat-devel packages),
       b2) local host is not a native Cray system: the conditionally built
           parts (basil_interface.c, libalps.la) are not built;
 3. updated configure logic:
    - since Cray support depends on mySQL, reordered tests in configure.ac,
    - reordered logic with regard to changes in (2),
    - an AM_CONDITIONAL to build native-Cray parts conditionally,
    - updated configure messages (XT/XE);
 4. run-time read_conf test to ensure use of select/cray is properly supported,
 5. an update of the NEWS file due to the change in (1) ==> may have a conflict
    in case you have a locally-updated copy.

I have compile-tested the three possible scenarios in (2).

    03_Cray-BASIL-node-ranking.diff
select/cray: perform node ranking

This supplies the select function-pointer to request a reordering of nodes based
on the current Cray node ordering. 

The Cray node ordering is set internally via the ALPS_NIDORDER configuration 
variables that controls the way ALPS considers nodes.

This ordering in turn determines the order of nodes as the appear subsequently 
in the Inventory output. The present patch exploits this fact and uses an
auto-incrementing number to reflect the node ranking (counting is reversed 
since the parser returns the nodes in stack/LIFO order).

The node ranking is performed on slurmctld (re-)configuration, hence the tests
are more stringent: exit if Inventory fails (this condition is extremely rare)
and if no nodes are powered up (also a condition that can be cured by restarting
slurmctld only when the system is ready).

    03_node-reordering-NodeHostName.diff

    02_Cray-BASIL-node-attributes-and-coordinates.diff

    02_salloc-no-node-update.diff

    01_Cray-BASIL-basic-support.diff plus
    01_changes-from-first-revision-of-patch-01.diff

    eliminates some inappropriate error messages. 01_interactive-no-script.diff

    expected start time was too far in the future for the backfill scheduler
    to compute.

    values as "*" rather than 65534. Patch from Rod Schulz, BULL.

    the job's priority using scontrol's "update jobid=..." rather than its 
    "hold" or "holdu" commands.

Fixed issue when using a storage_accounting plugin directly without the slurmDBD updates weren't always sent correctly to the slurmctld, appears to OS dependent, reported by Fredrik Tegenfeldt.

    clusters over 1000 nodes. Users can enable this by selecting the "Name"
    tab. This change dramatically improves scalability of sview.

    it were removed from the job's allocation. Now only the tasks on those
    nodes are terminated.

Fix for checking QOS to override partition limits, previously if not using QOS some limits would be overlooked.

    running or suspended states). Patch from Martin Perry, BULL.

    SelectTypeParameters=CR_ONE_TASK_PER_CORE.

Infrastructure in place plus logic in select/linear.

    indicate job termination was due to preemption.

    "*** JOB 65547 CANCELLED AT 2011-01-21T12:59:33 DUE TO PREEMPTION ***".

Added flag "NoReserve" to a QOS to make it so all jobs are created equal within a QOS.  So if larger, higher priority jobs are unable to run they don't prevent smaller jobs from running even if running the smaller jobs delay the start of the larger, higher priority jobs.

BLUEGENE - fixed issue where jobs wouldn't wait log enough for blocks to free and wanted to use blocks that are being freed for other jobs

-- BLUEGENE - This fixes a race condition with dynamic mode to make it so we copy the booted and job list of blocks before trying to create.

    so the highest job priority becomes TOP_PRIORITY.

    at FirstJobId.