NTHU-LSALAB / KubeShare

Share GPU between Pods in Kubernetes
Apache License 2.0
193 stars 42 forks source link

symbol lookup error: /kubeshare/library/libgemhook.so.1 #26

Open fung-hei-man opened 1 year ago

fung-hei-man commented 1 year ago

Hello!

I have installed KubeShare to my K8s cluster, tested with a simple pod specification.

apiVersion: v1
kind: Pod
metadata:
  name: test-kubeshare
    namespace: core
  labels:
    "sharedgpu/gpu_request": "0.1"
    "sharedgpu/gpu_limit": "0.2"
spec:
  schedulerName: kubeshare-scheduler
  containers:
    - name: pytorch
      image:  riyazhu/mnist:20220420
      imagePullPolicy: IfNotPresent

The result is positive.

GPU 0: NVIDIA GeForce RTX 3050 Ti Laptop GPU (UUID: GPU-xxx)

However, when I tried to deploy my own Docker image, following error occurs:

/bin/bash: symbol lookup error: /kubeshare/library/libgemhook.so.1: undefined symbol: __libc_dlopen_mode, version GLIBC_PRIVATE

Not sure if this information helps but my base image is nvidia/cuda:12.1.1-cudnn8-devel-ubuntu22.04 and I am compiling OpenCV for object detection using GPU. The image works fine when using the default-schduler.

Any suggestions on this issue? Thank you!