From 2023-01-18 ECMWF will enforce killing jobs if they have reached their wall time if "#SBATCH --time=" or command line option "-–time=" were provided with the job.
Alternatively ECMWF accepts jobs without "#SBATCH --time=" or command line option "-–time=" and ECMWF will instead use average runtimes of previous "similar" jobs and allow some grace before it will be killed.
"similar" jobs can be identified by a jobtag generated based on user ID, job name, job geometry and job output path.
After the first successful run the estimated runtime will be the average runtime of previous successful runs + one standard deviation.
For jobs were no recent previous runtime could be found, the assumed runtime is 24h plus an allowance of another 24h as grace time, allowing a new job to run for up to 48h.
As on the ECMWF Cray HPC the job will be given a 24h grace time to allow for system sessions and other issues holding up job progressing.
Example output which can be found in job output:
[ECMWF-INFO -sbatch] - ----------------------- [ECMWF-INFO -sbatch] - jobscript: /home/sycw/slurm/time.job [ECMWF-INFO -sbatch] - --- SCRIPT OPTIONS --- [ECMWF-INFO -sbatch] - #SBATCH --job-name=test [ECMWF-INFO -sbatch] - #SBATCH --nodes=2 [ECMWF-INFO -sbatch] - #SBATCH --mem-per-cpu=100 [ECMWF-INFO -sbatch] - #SBATCH --qos=np [ECMWF-INFO -sbatch] - #SBATCH --output=/home/sycw/slurm/time_test.out [ECMWF-INFO -sbatch] - ----------------------- [ECMWF-INFO -sbatch] - --- POST-PROCESSED OPTIONS --- [ECMWF-INFO -sbatch] - ARG --positional=['time.job'] [ECMWF-INFO -sbatch] - ARG --job_name=test [ECMWF-INFO -sbatch] - ARG --nodes=2 [ECMWF-INFO -sbatch] - ARG --output=/home/sycw/slurm/time_test.out [ECMWF-INFO -sbatch] - ARG --qos=np [ECMWF-INFO -sbatch] - ARG --mem_per_cpu=100 [ECMWF-INFO -sbatch] - ------------------------------ [ECMWF-INFO -sbatch] - jobtag: sycw-test-2x512-/home/sycw/slurm/time_test.out [ECMWF-INFO -sbatch] - ------------------------------ [ECMWF-INFO -sbatch] - Average Walltime 6 with a Standard Deviation 3 [ECMWF-INFO -sbatch] - Runtime history [ECMWF-INFO -sbatch] - Date | Cores Cluster Walltime Mem [ECMWF-INFO -sbatch] - 10.01.2023 - 12:18 | 512 ad 6 200M [ECMWF-INFO -sbatch] - 10.01.2023 - 12:17 | 512 ad 7 200M [ECMWF-INFO -sbatch] - ['/usr/bin/sbatch', '--job-name=test', '--nodes=2', '--output=/home/sycw/slurm/time_test.out', '--qos=np', '--mem-per-cpu=100', '--licenses=h2resw01', '--time=00:00:09', '/home/sycw/slurm/time.job'] OR [ECMWF-INFO -sbatch] - ------------------------------ [ECMWF-INFO -sbatch] - jobtag: sycw-testit6-1x2-/home/sycw/slurm/slurm-_JOBID_.out [ECMWF-INFO -sbatch] - ------------------------------ [ECMWF-INFO -sbatch] - Average Walltime 6 with a Standard Deviation 3 [ECMWF-INFO -sbatch] - Runtime history [ECMWF-INFO -sbatch] - Date | Cores Cluster Walltime Mem [ECMWF-INFO -sbatch] - 03.01.2023 - 15:24 | 2 ad 9 120000M [ECMWF-INFO -sbatch] - 12.12.2022 - 12:50 | 2 aa 6 1000M [ECMWF-INFO -sbatch] - 06.12.2022 - 15:29 | 2 ad 7 1000M [ECMWF-INFO -sbatch] - 05.12.2022 - 08:31 | 2 ac 7 1000M [ECMWF-INFO -sbatch] - 02.12.2022 - 10:15 | 2 aa 6 1000M [ECMWF-INFO -sbatch] - 02.12.2022 - 10:15 | 2 aa 6 1000M [ECMWF-INFO -sbatch] - 01.12.2022 - 11:00 | 2 aa 7 1000M [ECMWF-INFO -sbatch] - 30.11.2022 - 14:33 | 2 ad 7 1000M [ECMWF-INFO -sbatch] - 23.11.2022 - 13:34 | 2 ac 7 1000M [ECMWF-INFO -sbatch] - 23.11.2022 - 13:30 | 2 ac 7 1000M [ECMWF-INFO -sbatch] - 23.11.2022 - 13:23 | 2 ac 7 1000M [ECMWF-INFO -sbatch] - 23.11.2022 - 13:17 | 2 ac 6 1000M [ECMWF-INFO -sbatch] - 23.11.2022 - 13:14 | 2 ac 7 1000M [ECMWF-INFO -sbatch] - 23.11.2022 - 13:10 | 2 ac 7 1000M [ECMWF-INFO -sbatch] - ['/usr/bin/sbatch', '--cpus-per-task=1', '--job-name=testit6', '--ntasks=1', '--no-requeue', '--nodes=1', '--qos=nf', '--mem=120000m', '--licenses=h2resw01', '--time=00:00:09', '/home/sycw/slurm/gpil.job']