- Q: Why does my program not see any GPU device in the cluster? How do I specify the number of GPUs I want to use?
$ srun -p titanx-short --gres=gpu:1 your_script.sh your_argument_1 your_argument_2 ...
- Ans: You have to request from the scheduler to allocate GPU resources for your program. Use the parameter --gres-gpu:N in srun/sbatch to execute your program, where N is the number of desired GPUs at a compute node. All compute nodes currently have 4 GPUs, so at maximum you can use --gres-gpu:4. You also need to specify whether you will use titan/m40 GPUs (-p parameter) and the corresponding short/long queue. Example:
- Q: What is the faculty group I am assigned to?
$ id -g -nIf you are accidentally assigned to the wrong group, please email firstname.lastname@example.org and we will correct it.
- Ans: You can check the group you have been assigned to by typing:
- Q: How can I check my quota?
$ umquotaNote that /dev/sda7 refers to your home directory. The work1 and scratch1 have compression turned on. This can be a little confusing because it is not always obvious how much space your data is using. Some tools report the amount of space the compressed data takes up on disk rather than the size of the data itself. To see how much actual data there is, use du with the --apparent-size switch:$ du --apparent-size .
- Ans: The best way at the moment is to type the following from the head node:
- Q: How do I execute matlab on the cluster?
To compile your matlab program, type the following commands:$ srun --pty -L matlab@slurmdb -p matlab bash$ module load matlab/r2016a$ matlab -nojvm -r "mcc -mv test.m -d . -o test; exit"(here we assume that your matlab program entry is the test.m script)$ exit$ srun run_test.sh /cm/shared/apps/MATLAB/r2016a/(this will execute your program in the first available node - see also the rest of the FAQ on how to specify GPU resources)You may also add any program arguments e.g.:$ srun run_test.sh /cm/shared/apps/MATLAB/r2016a/ your_argument_1 your_argument_2Please also read the following resources on compiling matlab scripts for deployment:
- Ans: MATLAB R2016a is available on node100. It is the only node that can currently run MATLAB on gypsum. We do not recommend you to launch the matlab interface to execute your program because if every user does so, this node will be overloaded. Instead you should *compile* your matlab script into a deployable executable that you can launch on any other node of the cluster. There is one more advantage of using deployable executables: there are no license restrictions for the matlab's runtime libraries.
- Q: If some package is missing, how can I make it available?
- Ans: Please send a request to email@example.com and we will consider it asap.
- Q: I accidentally deleted some file(s). Could I restore them somehow?
- Ans: The work1 directory (/mnt/nfs/work1) is backed up regularly. You may retrieve previous snapshots of your files here: /mnt/nfs/work1/.zfs/snapshots