This website is deprecated, the old pages are kept online but you should refer in priority to the new web site hpc.uni.lu and the new technical documentation site hpc-docs.uni.lu
Which System to Use, According to My Workflow Patterns?
This is a recurring question, especially in light of the heterogeneity of clusters and their overlapping capabilities.
Here are the relevant guidelines, with cases and answers:
Nature of HPC activity | System(s) | Comments |
---|---|---|
For running either serial or parallel jobs | chaos/gaia | reserve exactly CPU resources you need |
For developing or trying out applications | chaos/gaia | reserve exactly CPU resources you need |
For small serial jobs | chaos/gaia | |
If you need linear scaling across nodes | chaos/gaia | Prefer whole-node reservations for #N>12 |
If you are running job-arrays | chaos/gaia | |
Use checkpointing to overcome X-days limit per job | chaos/gaia | Either CRIU or BLCR or user-space technologies should work |
For lengthy jobs, >3 days | chaos/gaia | ask a project description and explain operational aspects |
To run MATLAB | chaos/gaia | module load MATLAB |
To run serial Matlab jobs | chaos/gaia | remember to scale it with compiled code |
To run parallel Matlab jobs | chaos/gaia | remember to scale it with compiled code |
To run statistical computing software like R, Rmpi, Snow, Snowfall | chaos/gaia | The non-bare version of R in modules has strong features |
For post-processing and visualization | chaos/gaia | Normal system nodes should be used for this, ask for 1 core |
Run precompiled applications | chaos/gaia | long list is visible at [1], [2] |
Use EasyBuild to develop software | chaos/gaia | get familiar with the content at [3] |
If you need a large, single memory system | chaos/gaia | class bigmem offers 1TB per 32 cores (gaia-74/r-cluster-1) |
If you need a large, single memory system with AVX commandset | gaia | class bigmem offers 1TB per 32 cores (gaia-74) |
If you need a large, single memory system with multiple cores | gaia | node #73 is a bigsmp (ccNUMA) architecture with 160 cores |
If you need to experiment with accelerators and GPU computing | gaia | nodes #61-#72 have nVidia GPU cards; see. http://hpc.uni.lu |
For tightly coupled multinode parallel jobs w. low latency network | gaia | all nodes are on infiniband, although #74 is out of lustre |
To run Stata | gaia | module load Stata |
If you prefer Intel CPUs | chaos/gaia | Intel Cluster Studio XE 2013 has been purchased, use it |
If you prefer AMD CPUs | N/A |
The users are kindly asked to check what are the actual requirements for each application and use the resources appropriately and in a rational way.
[1] http://hpcbios.readthedocs.org/en/latest/HPCBIOS_2012-92.html
[2] https://github.com/hpcugent/easybuild/wiki/List-of-supported-software-packages