Parallel Computing with GPU

Development

Add a new task

To add a new task, create a task folder in src directory

mkdir -p mytask

Inside the folder, create a main.cu file. This will be the entry point of the task.

nano main.cu

Add a task util

To include a specific utils for a task, create a utils folder inside the task's folder

mkdir -p mytask/utils

Inside the folder, create a utils file

nano helper.h
nano helper.c

And include the header file to the main.cu file

extern "C"
{
#include "utils/helper.h"
}

Global utils

There is also src/utils directory which is the global utils folder. The code inside this directory will be compiled for all tasks.

Compiling the tasks

The tasks can be built using the make command

make

Running the tasks

The compiled tasks will be available in out directory as <taskname>.o. To run mytask.o, use the following command:

./out/mytask.o

Deployment

DGX-1 Cluster

Preparation

Create a pod file, for example pods.yml

nano pods.yml

with the following configuration:

apiVersion: v1
kind: Pod
metadata:
  # Pod Name
  name: user05-nvhpc-cuda
spec:
  restartPolicy: Never
  volumes:
      # Volume Name, should match spec.containers.volumeMounts.name
    - name: user05-nvhpc-cuda-pv
      persistentVolumeClaim:
        claimName: pemrograman-paralel-pv-claim
  containers:
    - name: app
      # Container Image
      image: "nvcr.io/nvidia/nvhpc:24.9-devel-cuda_multi-ubuntu22.04"
      command: ["/bin/sh"]
      args: ["-c", "while true; do echo 'pemrograman-paralel-pod log'; sleep 10; done"]
      volumeMounts:
          # Volume Name, should match spec.volumes.name
        - name: user05-nvhpc-cuda-pv
          mountPath: "/workspace"

or use environments/dgx-1/pods.yml

Deployment

Run the pod with the following command:

kubectl apply -f ./pods.yml

We assume that the kubectl's authentication is already set in the default config ~/.kube/config to access the cluster. You can also set --kubeconfig flag to use a custom config file.

Monitoring

Observe the pod creation with the following command:

kubectl describe pods user05-nvhpc-cuda

The pod's name user05-nvhpc-cuda in this example should match the pod's name that is deployed via pods.yml earlier (see metadata.name).

What we want to monitor is the pod's events. We should make sure that the pod creation process is running as expected.

Furthermore, check the pod's status using the following command:

kubectl get pods

If everything went well, the pod's status should be 'running'.

Usage

We can use the pod by exec-ing to the pod:

kubectl exec -it user05-nvhpc-cuda -- /bin/bash

After that, we can do whatever we want inside the pod. For example:

Check GPU information

nvidia-smi

Compile & run a CUDA program

nvcc program.cu -o program.o
./program.o

Post Usage

We can do cleanup to pods that are no longer used by deleting it using the following command:

kubectl delete pods user05-nvhpc-cuda