job_submit/lua - add "core_spec" to job_desc HPC-12614#79
Draft
iddecker wants to merge 76 commits intohpcugent:25.11.ugfrom
Draft
job_submit/lua - add "core_spec" to job_desc HPC-12614#79iddecker wants to merge 76 commits intohpcugent:25.11.ugfrom
iddecker wants to merge 76 commits intohpcugent:25.11.ugfrom
Conversation
When validating a node config, if there is some problem with the configuration, the node could end up with topo_cnt greater than the actual gres configured. This is a problem, because could lead to a corrupt state save being written by the controller: A topo_cnt sized array with less than topo_cnt values initialized. Which can cause SIGEGV while accessing some individual values. Fix properly updates topo_cnt when there are config errors to prevent that potential mismatch. Ticket: 24025 Changelog: Prevent potential controller segfault when reconfiguring after gres file updates. Cherry-picked: 0cb29ac
See merge request SchedMD/dev/slurm!2925
See merge request SchedMD/dev/slurm!2924
Ticket: 23467 Cherry-picked: 440cc02
When starting slurmd as a service, move the process to a subcgroup of the slurmd systemd unit cgroup. According to systemd docs the top cgroup is owned by systemd and it will eventually try to reset slurmd limits, e.g. on a systemctl daemon-reload. This can cause systemd errors in the system log, so we are only allowed to touch the limits in a sub-cgroup. This emulates DelegateSubgroup=slurmd only available after systemd 255. Ticket: 23467 Cherry-picked: 212bbf0
Ticket: 23467 Changelog: Reparent slurmd to a subcgroup to avoid conflicting with systemd. Cherry-picked: cb62c2f
See merge request SchedMD/dev/slurm!2929
sprio was unable to parse a comma-separated list of jobs despite documentation stating that this was possible. The entire job list was passed to the unfmt_job_id_string() function, but this function is unable to parse job lists. Switching to only pass the single job to unfmt_job_id_string() fixes this issue. Regression introduced in 4f21612. Changelog: Fix sprio regression not handling comma separated list of jobids. Ticket: 24625 Cherry-picked: 80beef6
See merge request SchedMD/dev/slurm!2950
Changelog: slurmctld,slurmd - Fix memory leak when container ID is populated. Issue: 50190 Cherry-picked: 13aba98
Enhanced _remove_ecores() to use a two-pass approach that collects all P-core frequencies instead of just the first one found. This fixes P-core detection on processors where different P-cores report different FrequencyMaxMHz values. The previous implementation (commit a80d6ba) used single-frequency matching which incorrectly excluded P-cores with different frequencies. On Intel Core Ultra 7 268V, some P-cores report 5000 MHz while others report 4900 MHz, causing only 2 of 4 P-cores to be detected. Implementation: - First pass: Find all CPU Kinds with CoreType=IntelCore and store all their distinct FrequencyMaxMHz values - Second pass: Include any CPU Kinds matching any collected frequency, even without CoreType=IntelCore attribute The second pass is necessary for hwloc < 2.10 where P-cores restricted by cpuset lose their CoreType attribute but retain FrequencyMaxMHz. This was fixed in hwloc 2.10 (commit 971ea80f9) which added PMU-based CoreType detection. The two-pass approach maintains compatibility with older hwloc versions while fixing the multiple-frequency limitation that broke detection on newer Intel processors. Changelog: slurmd - Fix P-core detection on processors with varying P-core frequencies and in cpuset-restricted environments. Ticket: 24590 Cherry-picked: c5bd248
See merge request SchedMD/dev/slurm!2957
See merge request SchedMD/dev/slurm!2955
This option allows namespace/linux to continue to operate even if bpf tokens are not supported on the system, user namespaces are enabled, and the cgroup constrain devices option is enabled. In this mode of operation, any devices that would be constrained will only be constrained to the job and not individual steps. Ticket: 21718 Changelog: namespace/linux - add disable_bpf_token option. Cherry-picked: a70b2d8
Ticket: 24652 Cherry-picked: 82ca853
Cherry-picked: c239700
Changelog: slurmctld - Avoid expedited requeue triggering a job to requeue when job exit code was zero. Ticket: 24564 Cherry-picked: c255c7d
batch_requeue_fini() should only be called after the epilog scripts have been complete. Add logic to catch and log if batch_requeue_fini() is ever called before the epilog is complete. Changelog: slurmctld - Avoid expedited requeue of jobs while waiting for job epilog script to complete. Ticket: 24564 Cherry-picked: 1511a97
Fair-Share factor reported only for users. Ticket: 24618 Cherry-picked: f0ce621
…nfig Prevent cloud nodes that are configured in topology.[conf|yaml] from being removed from the topology when the node is powered down if it does not use Topology=... in its node configuration line in slurm.conf or in the --conf slurmd option. If node_ptr->config_ptr->topology_str, node_ptr->topology_str, and node_ptr->topology_orig_str are NULL, then the slurm.conf and --conf option did not specify a topology for the node and thus the topology of the node does not need to be reset by node_mgr_set_node_topology(). Changelog: slurmctld - Prevent removing cloud nodes from the topology when putting them in the POWERED_DOWN state if they are present in topology.conf or topology.yaml and their node configuration did not specify the Topology option. Ticket: 24405 Cherry-picked: b36d9b4
Before topology_g_add_rm_node would only clear a node from a topology if topology_str was NULL. When it wasn't it would only modify the topologies listed in topology_str while potentially leaving the node in another topology. This makes it remove the node from all topologies not listed in topology_str. Changelog: interfaces/topology - When modifying a nodes topology with the Topology option in slurm.conf or the slurmd --conf Topology, change the topology to fully match the new topology. Ticket: 24405 Cherry-picked: fafc23e
If the node configuration line in slurm.conf specifies the Topology option, this change allows modifications to that option take effect on a reconfig/restart. The same for changes to Topology.[conf|yaml]. Changelog: slurmctld - Allow changes to topology.conf or topology.yaml, and slurm.conf node configuration Topology option to take effect on a reconfigure or restart when power saving is enabled. Ticket: 24405 Cherry-picked: 8f1b2d2
See merge request SchedMD/dev/slurm!2963
See merge request SchedMD/dev/slurm!2937
See merge request SchedMD/dev/slurm!2961
This fixes the logic in slurm_bf_licenses_equal() to validate that both license lists are equal. Before only licenses in the first license list were being checked to see if they were also the same in the second license list. It did not verify all licenses from the second license list were in the first license list. Changelog: slurmctld - Prevent backfill from combining future timeslots if they have different license reservations. Cherry-picked: fd0ef0d
If a slurmctld reconfig/restart happens while a cloud node is POWER_DOWN, the node's addr is reset, and, because it is CLOUD, it can end up FUTURE. Add POWER_DOWN to the other POWER states to exclude the node from addr reset. Ticket: 24358 Changelog: Fix CLOUD nodes infrequently becoming FUTURE on slurmctld restart. Cherry-picked: 49e3ba7
See merge request SchedMD/dev/slurm!2968
When rem_nodes is 0 in _get_block_level(), log2 resulted in -inf. Previously, this caused a crash. Check for negative before accessing the block_levels bitstring. Changelog: Fix handling of 0 node test allocations in topology/block. Ticket: 24552 Cherry-picked: 0b5f64f
See merge request SchedMD/dev/slurm!2973
See merge request SchedMD/dev/slurm!2962
Function returns the count of licenses in cluster_license_list. This is in preparation for following commit. Ticket: 24594 Cherry-picked: 95659c0
Sort the appended licenses from the advanced license reservations such that they come after the non-resv licenses, sorted by resv_id and lic_id. This is in preparation for the following commit. Ticket: 24594 Cherry-picked: aaa02a4
If the remaining count of a HRes license that the job requests changes in the next node_space table entry set later_start. Skip doing so in all other cases in regards to licenses. This is due to HRes licenses masking out nodes in the available bitmap depending on remaining counts. This prevents setting later_start unnecessarily leading to many unneeded calls to _try_sched(). Changelog: slurmctld - In backfill, prevent unnecessarily testing jobs at future times using the select plugin if it is guaranteed to fail. Ticket: 24594 Cherry-picked: a23a0a2
See merge request SchedMD/dev/slurm!2981
See merge request SchedMD/dev/slurm!2977
See merge request SchedMD/dev/slurm!2984
Update slurm.spec and debian/changelog as well.
slurmrestd does not handle SIGHUP; sending HUP terminates the daemon instead of reconfiguring. Drop ExecReload so systemctl reload is not advertised and does not kill the process. Changelog: slurmrestd - Remove ExecReload from unit file since the daemon does not handle SIGHUP (reload would terminate the process). Ticket: 24667 Cherry-picked: cd319ef
See merge request SchedMD/dev/slurm!2996
…es it Call _archive_table() for the main purge_type before archiving PURGE_JOB associated data (job env, script, etc.). Commit 4dcca88 had reversed this order; period_start is only set by the main archive path, while archive_write_file at the end of _archive_table uses it. The job env/ script codepaths (_pack_archive_job_env, _pack_archive_job_script) do not set it, so the main archive must run first. Changelog: Prevent "period_start should already be set" errors when purging slurmdbd data and fix file names for archives of purged slurmdbd data. Ticket: 24523 Cherry-picked: b263ac6
See merge request SchedMD/dev/slurm!3002
Ticket: 24439 Changelog: Skip x11 shutdown when x11 functionality was not requested. Cherry-picked: 165fe40
See merge request SchedMD/dev/slurm!3011
Changelog: Fix build errors with recent versions of libcurl (8.16+). Cherry-picked: 0ee19c3
See merge request SchedMD/dev/slurm!3018
Ticket: 24676 Cherry-picked: 372543c
See merge request SchedMD/dev/slurm!3021
After swapping alloc->environment with state.job_env in _alloc_job(), alloc->env_size was left with its original value while alloc->environment might be smaller, causing a segfault when trying to access it out-of-bounds in the later env_array_for_job() call. This failed with step_mgr and in 25.11 after 87e4a70. Both of two injects environment variables. Changelog: Fix scrun segfault with step_mgr and if environment is set. Ticket: 24662 Cherry-picked: 3ded4ee
See merge request SchedMD/dev/slurm!3028
Located in the job info struct. Ticket: 24674 Changelog: Fix two memory leaks located in the job info struct. Cherry-picked: 3012ffe
See merge request SchedMD/dev/slurm!3039
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
19 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.