1. Very High Cluster Utilization: The HPC cluster CARL and EDDY are very busy as many of you probably have noticed. On average only 8% of the available resources have been idle in the first four month of 2020 (and only 5% idle resources in April). Some users have reported very long queueing times, in particular for parallel jobs requesting multiple cores on a single node. In principle, the job scheduler should take care of the fair-sharing of the HPC resources, but in the end, there are only so many nodes. The following guidelines might help to improve the situation for everyone:
-
- If possible, limit the number of jobs you are running, if your jobs need more than the default memory and run for many days. For job arrays, use the maximum tasks limit.
- MPI Parallel jobs should best be submitted with the options --nodes and --ntasks-per-node to minimize the number of used nodes.
- Many single-core jobs can be organized to run on a limited number of nodes using the parallel command as explained in this HowTo (Section 4).
If you have questions, please contact hpcsupport@uol.de.
2. Changes Job Scheduler: About two weeks ago we have changed some of the settings of the job scheduler that determine job priorities. Mainly, we increased the time period for considering past usage so that less active users get higher priorities for their jobs. In addition, we also reserved a total of 28 compute nodes in the partition carl.p. This should decrease the wait times for (parallel) jobs requesting a time limit of 23:55h or less.
3. HPC Software Environments: Software on the cluster is organized in modules but in order to make sure different modules are compatible with each other we also use so-called HPC software environments. These environments can be activated by loading one of the hpc-env modules. Most software is installed in two environments: hpc-uniol-env and hpc-env/6.4 and we now started installing software in hpc-env/8.3 (see software news below). The new environment includes more recent compilers and toolchains. If you are looking for a certain software package, you can use module spider to find out which environment you need to load. If you are missing software in one of the environments, please contact hpcsupport@uol.de.
4. Dates: Please take note following upcoming events
-
- Workshop Supercomputing@UOL: postponed, no new date set yet
- Meeting of the HPC user representatives: postponed, no new date set yet
5. Software News: The following modules are now available on hpc-env/8.4:
- TensorFlow/2.1.0-foss-2019b-Python-3.7.4
- picard/2.22.3-Java-11
- GATK/4.1.4.1-GCCcore-8.3.0-Java-11
- BCFtools/1.9-foss-2019b.eb
- ConnectomeWorkbench/1.3.2-GCCcore-8.3.0.eb
- Bowtie2/2.4.1-GCC-8.3.0
- HDF5/1.10.5-gompic-2019b
- HDF5/1.10.5-gompi-2019b
- GATK/4.1.4.1-GCCcore-8.3.0-Java-11
- picard/2.22.3-Java-11
- MIRTK/2020.04.foss/2019b
The list above might not be complete, you can always search for software with the command “module spider <softwarename>”.