Ticket 5821

Hi

Could you send me the output from this gdb commands?
t 1 
f 5
p *job_ptr

Dominik

(gdb) t 1
[Switching to thread 1 (Thread 0x7f0d69234740 (LWP 21567))]
#0  0x00007f0d68cdef38 in _job_alloc (core_bitmap=0x2c93b40, node_name=0x28f4790 "nvl-002", job_id=290382, 
    gres_name=0x229ccb0 "gpu", node_offset=0, node_cnt=1, node_gres_data=0x2a2ef20, job_gres_data=0x2b0fa70)
    at gres.c:5856
5856	in gres.c
(gdb) f 5
#5  0x000000000044cef8 in reset_job_bitmaps () at job_mgr.c:10919
10919	job_mgr.c: No such file or directory.
(gdb) p *job_ptr
$1 = {magic = 4038539564, account = 0x2b0f6d0 "general", admin_comment = 0x0, alias_list = 0x0, 
  alloc_node = 0x2b0f6a0 "dev-intel18", alloc_resp_port = 52491, alloc_sid = 29888, array_job_id = 0, 
  array_task_id = 4294967294, array_recs = 0x0, assoc_id = 2762, assoc_ptr = 0x2403620, batch_features = 0x0, 
  batch_flag = 0, batch_host = 0x2ac7f80 "nvl-002", billable_tres = 8, bit_flags = 32768, burst_buffer = 0x0, 
  burst_buffer_state = 0x0, check_job = 0x0, ckpt_interval = 0, ckpt_time = 0, clusters = 0x0, 
  comment = 0x2b0f6f0 "stdout=/mnt/home/dunan/slurm-290382.out", cpu_cnt = 8, cpus_per_tres = 0x0, cr_enabled = 0, 
  db_index = 2958410, deadline = 0, delay_boot = 0, derived_ec = 0, details = 0x2b0f470, direct_set_prio = 0, 
  end_time = 1539559241, end_time_exp = 1539559241, epilog_running = false, exit_code = 0, fed_details = 0x0, 
  front_end_ptr = 0x0, gids = 0x0, gres_list = 0x2804ec0, gres_alloc = 0x2b0f750 "7696487:1", gres_detail_cnt = 1, 
  gres_detail_str = 0x2b114f0, gres_req = 0x2b0f780 "gpu:1", gres_used = 0x0, group_id = 2003, job_id = 290382, 
  job_next = 0x0, job_array_next_j = 0x0, job_array_next_t = 0x0, job_resrcs = 0x2b0f7f0, job_state = 1, 
  kill_on_node_fail = 1, last_sched_eval = 1538954441, licenses = 0x0, license_list = 0x0, limit_set = {qos = 0, 
    time = 0, tres = 0x2b0efe0}, mail_type = 0, mail_user = 0x0, mem_per_tres = 0x0, mcs_label = 0x0, 
  name = 0x2b0efc0 "sh", network = 0x0, next_step_id = 1, ngids = 0, nodes = 0x2b0ef50 "nvl-002", 
  node_addr = 0x2c85ea0, node_bitmap = 0x2c85bd0, node_bitmap_cg = 0x0, node_cnt = 1, node_cnt_wag = 0, 
  nodes_completing = 0x0, origin_cluster = 0x2b11290 "msuhpcc", other_port = 52490, pack_job_id = 0, 
  pack_job_id_set = 0x0, pack_job_offset = 0, pack_job_list = 0x0, 
  partition = 0x2b0ef70 "general-long-18,general-long-14,general-long-16", part_ptr_list = 0x26cf950, 
  part_nodes_missing = false, part_ptr = 0x28f5e90, power_flags = 0 '\000', pre_sus_time = 0, preempt_time = 0, 
  preempt_in_progress = false, priority = 3389, priority_array = 0x0, prio_factors = 0x2b0ee90, profile = 0, 
  qos_id = 1, qos_ptr = 0x22c17a0, qos_blocking_ptr = 0x0, reboot = 0 '\000', restart_cnt = 0, resize_time = 0, 
  resv_id = 0, resv_name = 0x0, resv_ptr = 0x0, requid = 4294967295, resp_host = 0x2b0ef20 "192.168.9.200", 
  sched_nodes = 0x0, select_jobinfo = 0x2b0f7c0, spank_job_env = 0x0, spank_job_env_size = 0, 
  start_protocol_ver = 8192, start_time = 1538954441, state_desc = 0x0, state_reason = 0, state_reason_prev = 0, 
  step_list = 0x26cf9a0, suspend_time = 0, system_comment = 0x0, time_last_active = 1539005708, time_limit = 10080, 
  time_min = 0, tot_sus_time = 0, total_cpus = 8, total_nodes = 1, tres_bind = 0x0, tres_freq = 0x0, 
  tres_per_job = 0x0, tres_per_node = 0x2b0f730 "gpu:1", tres_per_socket = 0x0, tres_per_task = 0x0, 
  tres_req_cnt = 0x2b11340, tres_req_str = 0x2b112b0 "1=8,2=6000,4=1,1001=1", 
  tres_fmt_req_str = 0x2b11250 "cpu=8,mem=6000M,node=1,gres/gpu=1", tres_alloc_cnt = 0x2b112e0, 
  tres_alloc_str = 0x2b113a0 "1=8,2=6000,3=18446744073709551614,4=1,5=8,1001=1", 
  tres_fmt_alloc_str = 0x2b113f0 "cpu=8,mem=6000M,node=1,billing=8,gres/gpu=1", user_id = 721898, 
  user_name = 0x2b0f680 "dunan", wait_all_nodes = 0, warn_flags = 0, warn_signal = 0, warn_time = 0, wckey = 0x0, 
  req_switch = 0, wait4switch = 0, best_switch = true, wait4switch_start = 0}

Dominik,

I reverted our production server back to 17.11.7 for now. I'm going to change this to a severity 2.

Hi

Good to know that you are "up" now.

Could you send also job_gres_ptr?
t 1
f 0
p *job_gres_ptr

Dominik

Hi

I can recreate this issue. I will let you know when I fix this.

Dominik

Hi

I have the preliminary patch now, but it still needs some review, let me know if you need it.
I suggest staying on 17.11 slurmctld until it will be officially in git repo.


Dominik

Dominik,

That's good news! We are aiming to be on 18.08.1 by next Monday. Do you think this internal review will be done by then?

Thanks,
Steve

Hi

If it wasn't US holiday, it would be done already. Because of Columbus Day, it will take a little bit longer, it should be in repo tomorrow.

Dominik

Hey Dominik,

I'm hoping to do this upgrade later this evening. Will this patch be available today?

Thanks,
Steve

Comment on attachment 7994 [details]
Patch to create structures outside of if

Steve, please use this patch it should get you what you need.  It will be in 18.08.2 which will probably be tagged in a few weeks commit 33f8f22d.

Danny,

We updated using this patch and slurmctld runs, however, we're seeing a lot of the following errors in the log:

[2018-10-10T21:28:54.125] error: gres/gpu: job 326801 dealloc node nvl-006 topo gres count underflow (0 4)
[2018-10-10T21:28:54.125] error: gres/gpu: job 326801 dealloc node nvl-006 type v100 gres count underflow (0 4)
[2018-10-10T21:28:54.127] error: gres/gpu: job 303352 dealloc node nvl-002 topo gres count underflow (0 1)
[2018-10-10T21:28:54.127] error: gres/gpu: job 303352 dealloc node nvl-002 type v100 gres count underflow (0 1)
[2018-10-10T21:28:54.127] error: gres/gpu: job 326800 dealloc node nvl-006 topo gres count underflow (1 4)
[2018-10-10T21:28:54.127] error: gres/gpu: job 326800 dealloc node nvl-006 type v100 gres count underflow (0 4)
[2018-10-10T21:28:54.127] error: gres/gpu: job 326801 dealloc node nvl-006 topo gres count underflow (0 4)
[2018-10-10T21:28:54.128] error: gres/gpu: job 326801 dealloc node nvl-006 type v100 gres count underflow (0 4)
[2018-10-10T21:28:54.155] error: gres/gpu: job 303352 dealloc node nvl-002 topo gres count underflow (0 1)
[2018-10-10T21:28:54.155] error: gres/gpu: job 303352 dealloc node nvl-002 type v100 gres count underflow (0 1)
[2018-10-10T21:28:54.157] error: gres/gpu: job 303352 dealloc node nvl-002 topo gres count underflow (0 1)
[2018-10-10T21:28:54.157] error: gres/gpu: job 303352 dealloc node nvl-002 type v100 gres count underflow (0 1)
[2018-10-10T21:28:54.160] error: gres/gpu: job 303352 dealloc node nvl-002 topo gres count underflow (0 1)
[2018-10-10T21:28:54.160] error: gres/gpu: job 303352 dealloc node nvl-002 type v100 gres count underflow (0 1)
[2018-10-10T21:28:54.174] error: gres/gpu: job 326800 dealloc node nvl-006 topo gres count underflow (1 4)
[2018-10-10T21:28:54.174] error: gres/gpu: job 326800 dealloc node nvl-006 type v100 gres count underflow (0 4)
[2018-10-10T21:28:54.174] error: gres/gpu: job 326801 dealloc node nvl-006 topo gres count underflow (0 4)
[2018-10-10T21:28:54.174] error: gres/gpu: job 326801 dealloc node nvl-006 type v100 gres count underflow (0 4)
[2018-10-10T21:28:54.177] error: gres/gpu: job 326800 dealloc node nvl-006 topo gres count underflow (1 4)

Hi

We expect this is only a transitional thing and it should go away after all 17.11 jobs are gone. I will inform you when I will find what exactly caused this problem.

Dominik

Hi

Could you send me full slurmctld.log
and output from "scontrol show job <one of affected job_id eg. 326800 >", if such jobs still exist in the system?

Dominik

# scontrol show job 326800
JobId=326800 JobName=4zh9_long_200ns_MD
   UserId=foouser(*) GroupId=hpccclass(*) MCS_label=N/A
   Priority=20007 Nice=0 Account=general QOS=normal
   JobState=RUNNING Reason=None Dependency=(null)
   Requeue=0 Restarts=0 BatchFlag=1 Reboot=0 ExitCode=0:0
   RunTime=1-14:46:46 TimeLimit=7-00:00:00 TimeMin=N/A
   SubmitTime=2018-10-10T18:39:43 EligibleTime=2018-10-10T18:39:43
   AccrueTime=Unknown
   StartTime=2018-10-10T18:39:43 EndTime=2018-10-17T18:39:43 Deadline=N/A
   PreemptTime=None SuspendTime=None SecsPreSuspend=0
   LastSchedEval=2018-10-10T18:39:43
   Partition=general-long-18 AllocNode:Sid=lac-249:2131
   ReqNodeList=(null) ExcNodeList=(null)
   NodeList=nvl-006
   BatchHost=nvl-006
   NumNodes=1 NumCPUs=1 NumTasks=1 CPUs/Task=1 ReqB:S:C:T=0:0:*:*
   TRES=cpu=1,mem=12G,node=1,billing=1,gres/gpu=4
   Socks/Node=* NtasksPerN:B:S:C=1:0:*:* CoreSpec=*
   MinCPUsNode=1 MinMemoryNode=12G MinTmpDiskNode=0
   Features=intel16|intel18 DelayBoot=00:00:00
   OverSubscribe=OK Contiguous=0 Licenses=(null) Network=(null)
   Command=/mnt/home/foouser/MD/CRBP/longRun/4zh9_long_200ns_MD.sb
   WorkDir=/mnt/home/foouser/MD/CRBP/longRun
   Comment=stdout=/mnt/home/foouser/MD/CRBP/longRun/4zh9_long_200ns_MD.sb.o%A 
   StdErr=/mnt/home/foouser/MD/CRBP/longRun/4zh9_long_200ns_MD.sb.e326800
   StdIn=/dev/null
   StdOut=/mnt/home/foouser/MD/CRBP/longRun/4zh9_long_200ns_MD.sb.o326800
   Power=
   TresPerNode=gpu:4

Hi

Does this error still occur?
I am still working on this but I can't reproduce it.

Dominik

Hi

Can we drop severity to 3 now, as the patch is already in git repo?

Dominik

Dominik,

This is fine to drop to sev 3.

I am still seeing this error in the logs. I'm now seeing it for jobs submitted after the update. For example, job 672084:

[2018-10-25T12:23:44.040] error: gres/gpu: job 672084 dealloc node lac-196 gres count underflow
[2018-10-25T12:23:44.040] error: gres/gpu: job 672084 dealloc node lac-196 topo gres count underflow (3 4)
[2018-10-25T12:23:44.040] error: gres/gpu: job 672084 dealloc node lac-196 type k80 gres count underflow (3 4)


JobId=672084 JobName=bash
   UserId=foouser(*) GroupId=group(*) MCS_label=N/A
   Priority=24787 Nice=0 Account=general QOS=normal
   JobState=FAILED Reason=NonZeroExitCode Dependency=(null)
   Requeue=0 Restarts=0 BatchFlag=0 Reboot=0 ExitCode=130:0
   RunTime=00:14:28 TimeLimit=00:20:00 TimeMin=N/A
   SubmitTime=2018-10-25T12:09:16 EligibleTime=2018-10-25T12:09:16
   StartTime=2018-10-25T12:09:16 EndTime=2018-10-25T12:23:44 Deadline=N/A
   PreemptTime=None SuspendTime=None SecsPreSuspend=0
   LastSchedEval=2018-10-25T12:09:16
   Partition=general-long-16 AllocNode:Sid=dev-intel16-k80:18489
   ReqNodeList=(null) ExcNodeList=(null)
   NodeList=lac-196
   BatchHost=lac-196
   NumNodes=1 NumCPUs=8 NumTasks=4 CPUs/Task=2 ReqB:S:C:T=0:0:*:*
   TRES=cpu=8,mem=2G,node=1,billing=8,gres/gpu=4
   Socks/Node=* NtasksPerN:B:S:C=0:0:*:* CoreSpec=*
   MinCPUsNode=2 MinMemoryNode=2G MinTmpDiskNode=0
   Features=gpgpu DelayBoot=00:00:00
   Gres=gpu:4 Reservation=(null)
   OverSubscribe=OK Contiguous=0 Licenses=(null) Network=(null)
   Command=/bin/bash
   WorkDir=/mnt/home/foouser/Documents/ShearWaveImaging/CUDA_GFS_ForHPCC_Release
   Comment=stdout=/mnt/home/foouser/Documents/ShearWaveImaging/CUDA_GFS_ForHPCC_Release/slurm-672084.out 
   Power=

Thanks,
Steve

Created attachment 8105 [details]
Slurmctld log

Hi

It seems that you can easily reproduce this.
Because I still can't reproduce this.
Could you generate slurmctld.log with enabled DebugFlags SelectType?
eg.:
scontrol setdebugflags +SelectType
submit job
wait for log
scontrol setdebugflags -SelectType

Could you also send me slurm.conf and exact command you used to submit this job?


Dominik

Dominik,

I have not been able to submit a job that produces these errors, I am only noticing these jobs after the fact. I'll attach the slurm log with the selecttype flag set. Perhaps the information from that flag for the jobs that are already throwing this error will be helpful.

Best,
Steve

Created attachment 8202 [details]
Slurmctld log +SelectType, compressed with xz

Hi

It looks like logic in select plugin is working well, a problem must be in a different place.
Could you generate slurmctld.log with enabled DebugFlags TraceJobs?

Dominik

Created attachment 9233 [details]
Slurmctld log +TraceJobs

Dominik,

Here is the log output with TraceJobs enabled.

Best,
Steve

Hi

This two commit should fix (gres count underflow) issue:
https://github.com/SchedMD/slurm/commit/6f8cd92e1091e
https://github.com/SchedMD/slurm/commit/69d78159c3305
Both patches will be included in 18.08.6.

In separate bug, we handle wrongly set debug level which returns a false positive message like this
"select_nodes: calling _get_req_features() for JobId=10129368 with not NULL job resources"

Let me know if we can close this bug now.

Dominik

Hi

I will close this ticket as resolved/fixed and set "Version Fixed" to 18.08.2 (this is version were an original bug was fixed).
Please open a new ticket if you notice more "count underflow" message above version 18.08.6.


Dominik