This is a GPU server of the Department of Computer Engineering, University of Peradeniya.
You can use this server using the LDAP login for ce.pdn.ac.lk undergraduates (which is used for CO top floor lab, aiken and tesla) without sudo access.
Short answer: You should inform the admin if you are using more than 4GB of GPU memory.
Long answer: You are allowed to use the whole GPU for your work. But there are some guidelines to make sure you don't block others from using the GPU unnecessarily. The Tesla K40C card on this server can be used by more than one person simultaneously in most situations.
GPU usage can be checked by running the command nvidia-smi. The CPU and RAM usage can be checked by running the command htop.
If a user is blocking the GPU by using more than 4GB at a given moment: Contact the admin and request him/her to kill the process.
If a user is blocking the GPU by using more than 4GB at a given moment: Contact the admin and request him/her to kill the process.
If you are using the GPU for deep learning: If your code is implemented using a neural network framework, it might try to allocate the full GPU memory even when it is not needed. This issue can be corrected by modifying the code.
This is the code to be added for tensorflow 2.2 or newer.
import tensorflow as tf
gpus = tf.config.experimental.list_physical_devices('GPU')
if gpus:
try:
for gpu in gpus:
tf.config.experimental.set_memory_growth(gpu, True)
except RuntimeError as e:
print(e)
You may find such code to correct this issue in similar frameworks as well.
If you make this change and your NN training still requires more than 4GB of GPU memory, you can try setting a smaller batch_size in the NN training step. Following is a piece of code for Keras.
nn.fit(xTrain,yTrain,epochs=100,verbose=1,batch_size=8)
If you cannot reduce the memory footprint of your NN training, please inform the admin so that s/he will not kill your process to give space to another user.
If you are using the GPU for something else: It is okay to use the full GPU memory for short periods of time (around 5 min). But if you use the GPU for a longer period of time you should inform the admin about your requirement.
We store frequently used datasets inside the kepler.ce.pdn.ac.lk local storage for easy access. You can access these by the following commands
ls /storage/datasets/
These datasets are in the read only mode. You can use these datsets as the input for your tasks but you have to output the results to your home directory. The following datasets are stored at the moment.
If you have suggestions for what datasets should be made available on this server, please send a request via email available in https://www.cs.umd.edu/~gihan/contact/ address.
ls /storage/datasets/
nvcc code.cu -arch=sm_35
The -arch=sm_35 flag is required to compile code properly on this server.