Current Allennlp Docker Image - has the userfunction as super user.

[mshanmugam]

Hi Team ,

Current Allenai docker image is having super user(sudo/as root) function which we cannot bring it inside our own environments. Is there a way we can we can disable them via any configurations / env variables.

[epwalsh]

Hey @mshanmugam, I'm not 100% sure what you mean. Are you talking about how you have to run Docker with sudo permissions?

[dirkgr]

Are you trying to run the AllenNLP docker image on a computer where you can't run Docker? An alternative way to go might be to install miniconda, and then install AllenNLP into the conda environment (with pip install allennlp). You'll have to learn a bit about conda though.

[mshanmugam]

Actually we are trying to use the docker image as is and our security scan says , allennlp docker image has some root privilege function so we couldn't pass the scan.

On the other note , we were able to manually install the allennlp on docker kubernetes platform , with 2 core cpu to the container and 4GB memory , we are seeing 10X slower performance on container compared to the baremetal/regular VM.
we don't use GPU , we use only CPU(No Cuda/GPU) on both containers and baremetal. The model that is being used is BiDAF with glove embedding for reading comprehension. Please let me know if need to open a separate discussion thread for the slowness / performance.

Thanks for your help.

[dirkgr]

Does it give any extra information? What root privilege function is it complaining about?

I would guess the performance issue is down to the version of torch you're installing. Is there a difference between these versions?

[mshanmugam]

For Root ,it doesnt say any specific commands , it just says the image needs to function as a super user.

For the performance the torch libs are same .. We are using 2.1.0 model , and torch version is 1.7.1 on python 3.6.
Only difference is baremetal/vm has 8 cores , the container is allocated with 2 cores. But I couldn't visibly see the cpu usage being high on containers,its using less than 100 milli cores.

[dirkgr]

There are different ways torch 1.7.1 could be built, even when it's just doing CPU. I tend to use the version from Conda all the time, but if you pip install, you will get a different version.