1
0
Fork 0
mirror of https://github.com/MillironX/nf-configs.git synced 2024-11-28 02:29:54 +00:00
nf-configs/docs/uppmax.md

67 lines
2.8 KiB
Markdown
Raw Normal View History

2019-01-07 13:25:06 +00:00
# nf-core/configs: UPPMAX Configuration
All nf-core pipelines have been successfully configured for use on the Swedish UPPMAX clusters.
## Using the UPPMAX config profile
2019-11-25 23:48:17 +00:00
2019-12-05 14:18:01 +00:00
To use, run the pipeline with `-profile uppmax` (one hyphen).
This will download and launch the [`uppmax.config`](../conf/uppmax.config) which has been pre-configured with a setup suitable for the UPPMAX servers.
Using this profile, a docker image containing all of the required software will be downloaded, and converted to a Singularity image before execution of the pipeline.
2019-01-07 13:25:06 +00:00
In addition to this config profile, you will also need to specify an UPPMAX project id.
You can do this with the `--project` flag (two hyphens) when launching nextflow. For example:
```bash
nextflow run nf-core/PIPELINE -profile uppmax --project SNIC 2018/1-234 # ..rest of pipeline flags
```
Before running the pipeline you will need to either install Nextflow or load it using the environment module system.
This config enables Nextflow to manage the pipeline jobs via the Slurm job scheduler.
Just run Nextflow on a login node and it will handle everything else.
## Using iGenomes references
2019-11-25 23:48:17 +00:00
2019-01-07 13:25:06 +00:00
A local copy of the iGenomes resource has been made available on all UPPMAX clusters so you should be able to run the pipeline against any reference available in the `igenomes.config`.
You can do this by simply using the `--genome <GENOME_ID>` parameter.
## Getting more memory
2019-11-25 23:48:17 +00:00
2019-01-07 13:25:06 +00:00
If your nf-core pipeline run is running out of memory, you can run on a fat node with more memory using the following nextflow flags:
```bash
--clusterOptions "-C mem256GB" --max_memory "256GB"
```
This raises the ceiling of available memory from the default of `128.GB` to `256.GB`.
Rackham has nodes with 128GB, 256GB and 1TB memory available.
Note that each job will still start with the same request as normal, but restarted attempts with larger requests will be able to request greater amounts of memory.
All jobs will be submitted to fat nodes using this method, so it's only for use in extreme circumstances.
2019-12-05 14:18:01 +00:00
## How to specify a UPPMAX cluster
The UPPMAX nf-core configuration profile uses the `hostname` of the active environment to automatically apply the following resource limits:
2019-12-05 14:18:01 +00:00
* `bianca`
* cpus available: 16 cpus
* memory available: 109 GB
* `irma`
* cpus available: 16 cpus
* memory available: 250 GB
* `rackham`
* cpus available: 20 cpus
* memory available: 125 GB
## Development config
2019-11-25 23:48:17 +00:00
2019-12-05 14:18:01 +00:00
If doing pipeline development work on UPPMAX, the `devel` profile allows for faster testing.
2019-01-07 13:25:06 +00:00
Applied after main UPPMAX config, it overwrites certain parts of the config and submits jobs to the `devcore` queue, which has much faster queue times.
All jobs are limited to 1 hour to be eligible for this queue and only one job allowed at a time.
It is not suitable for use with real data.
To use it, submit with `-profile uppmax,devel`.