Installing NVIDIA Driver + HPL + missing libcuda.so

Hello everybody,

We are trying to benchmark our cluster using LinPack benchmark tool; using NVIDIA-HPL.
We managed to install it successfully. However, we wanna run HPL from a Head Node, i.e. not locally.

When trying to build/run HPL on the Head Node, HPL complains at libcuda.so!! “Not found”

I know this library comes with NVIDIA Driver. Unfortunately, we couldn’t even install the driver because the Head Node doesn’t have an NVIDIA GPU device.

Is there any way to install the driver on a machine which doesn’t have an NVIDIA GPU OR any other packages which libcuda.so comes with??

Thanks for any help?
Husain

Hello everybody,

We are trying to benchmark our cluster using LinPack benchmark tool; using NVIDIA-HPL.
We managed to install it successfully. However, we wanna run HPL from a Head Node, i.e. not locally.

When trying to build/run HPL on the Head Node, HPL complains at libcuda.so!! “Not found”

I know this library comes with NVIDIA Driver. Unfortunately, we couldn’t even install the driver because the Head Node doesn’t have an NVIDIA GPU device.

Is there any way to install the driver on a machine which doesn’t have an NVIDIA GPU OR any other packages which libcuda.so comes with??

Thanks for any help?
Husain

I’m not at all familiar with the things you are using. Why don’t you just SSL to your machines that have nvidia GPUs and do the things on that machine directly?

Anyway, libcuda.so should be under /usr/lib (depending on how you installed the driver on your machines, you might have to use locate to find the libcuda.so sometimes)

I guess you could directly copy it and do the compilation on a machine without the driver.

Hi Husain, I believe that you can install the driver and toolkits on non default locations. I am referring to v4.0.17.

On another note, how did you find HPL + CUDA from Nvidia? I am most interested in this code. as I would like to benchmark our cluster with GPUs.

regards

Michael