diff --git a/.github/actions/spelling/allow.txt b/.github/actions/spelling/allow.txt index e96d734a..7094be37 100644 --- a/.github/actions/spelling/allow.txt +++ b/.github/actions/spelling/allow.txt @@ -16,6 +16,7 @@ CWP CXI Ceph Containerfile +DCGM DNS Dockerfiles Dufourspitze @@ -92,6 +93,7 @@ Piz Plesset Podladchikov Pulay +PyPi RCCL RDMA ROCm @@ -166,7 +168,9 @@ gpu gromos groundstate gsl +gssr hdf +heatmaps hotmail huggingface hwloc @@ -203,6 +207,7 @@ mkl mpi mps multitenancy +mycontainer nanoscale nanotron nccl @@ -241,6 +246,7 @@ kubeconfig ceph rwx rwo +sqsh subdomain tls kured diff --git a/docs/images/gssr/heatmap_eg.png b/docs/images/gssr/heatmap_eg.png new file mode 100644 index 00000000..dfd674ea Binary files /dev/null and b/docs/images/gssr/heatmap_eg.png differ diff --git a/docs/images/gssr/timeseries_eg.png b/docs/images/gssr/timeseries_eg.png new file mode 100644 index 00000000..37346f19 Binary files /dev/null and b/docs/images/gssr/timeseries_eg.png differ diff --git a/docs/software/gssr/containers.md b/docs/software/gssr/containers.md new file mode 100644 index 00000000..b841c58b --- /dev/null +++ b/docs/software/gssr/containers.md @@ -0,0 +1,120 @@ +[](){#ref-gssr-containers} +# gssr - Containers Guide + +CSCS highly recommends that all users leverage on container solutions on our Alps platforms so as to flexibly configure any required user environments of their choice within the containers. Users thus have maximum flexibility as they are not tied to any specific operating systems and/or software stacks. + +The following guide will explain how to install and use `gssr` within a container. + +Most CSCS users leverage on the base containers with pre-installed CUDA from Nvidia. As such, in the following documentation, we will use a PyTorch base container as an example. + +## Preparing a container with `gssr` + +### Base Container from Nvidia + +The most commonly used Nvidia container used on Alps is the [Nvidia's PyTorch container](https://catalog.ngc.nvidia.com/orgs/nvidia/containers/pytorch). Typically the latest version is preferred for the most up-to-date functionalities of PyTorch. + +#### Example: Preparing a Nvidia PyTorch ContainerFile +``` +FROM --platform=linux/arm64 nvcr.io/nvidia/pytorch:25.08-py3 + +ENV DEBIAN_FRONTEND=noninteractive + +RUN apt-get update \ + && apt-get install -y wget rsync rclone vim git htop nvtop nano \ + && apt-get clean \ + && rm -rf /var/lib/apt/lists/* + +# Installing gssr +RUN pip install gssr + +# Install your application and dependencies as required +... +``` +As you can see from the above example, gssr can easily be installed with a `RUN pip install gssr` command. + +Once your `ContainerFile` is ready, you can build it on any Alps platforms with the following commands to create a container with label `mycontainer`. + +```bash +srun -A {groupID} --pty bash +# Once you have an interactive session, use podman command to build your container +# -v is to mount the fast storage on Alps into the container. +podman build -v $SCRATCH:$SCRATCH -t mycontainer:0.1 . +# Export the container from the podman's cache to a local sqshfs file with enroot +enroot import -x mount -o mycontainer.sqsh podman://local:mycontainer:0.1 +``` + +Now you should have a sqsh file of your container. Please note that you should replace `mycontainer` label to any other label of your choice. The version `0.1` can also be omitted or replaced with another version as required. + +## Create CSCS configuration for Container + +The next step is to tell CSCS container engine solution where your container is and how you would like to run it. To do so, you will have to create a`{label}.toml` file in your `$HOME/.edf` directory. + +### Example of a `mycontainer.toml` file +``` +image = "/capstor/scratch/cscs/username/directoryWhereYourContainerIs/mycontainer.sqsh" +mounts = ["/capstor/scratch/cscs/username:/capstor/scratch/cscs/username"] +workdir = "/capstor/scratch/cscs/username" +writable = true + +[annotations] +com.hooks.dcgm.enabled = "true" +``` + +Please note that the `mounts` line is important if you want $SCRATCH to be available in your container. You can also mount a specific directory or file in $HOME and/or $SCRATCH as required. You should modify the username and the image directory as per your setup. + +To use `gssr` in a container, you will need the `dcgm` hook that is configured in the `[annotations]` section to enable DCGM libraries to be available within the container. + +### Run the application and container with gssr + +To invoke `gssr`, you can do the following in your sbatch file. + +#### Example of a mycontainer.sbatch file +``` +#!/bin/bash +#SBATCH -N4 +#SBATCH -A groupname +#SBATCH -J mycontainer +#SBATCH -t 1:00:00 +#SBATCH ... + +srun --environment=mycontainer bash -c 'gssr --wrap="python abc.py"' + +``` + +Please replace the text `...` for any other SBATCH configuration that your job requires. +The `--environment` flag tells Slurm which container (name of the toml file) you would like to run. +The `bash -c` requirement is to initialise the bash environment within your container. + +If no `gssr` is used, the `srun` command in your container should like that.: + +``` +srun --environment=mycontainer bash -c 'python abc.py'. +``` + +Now you are ready to submit your sbatch file to slurm with `sbatch` command. + +## Analyze the output + +Once your job successfully concluded, you should find a folder named `profile_out_{slurm_jobid}` where `gssr` json outputs are in. + +To analyze the outputs, you can do so interactively within any containers where `gssr` is installed, e.g., `mycontainer` we have in this guide. + +To get an interactive session of this container: + +``` +srun -A groupname --environment=mycontainer --pty bash +cd {directory where the gssr output data is generated} +``` +Alternatively, you can install `gssr` locally and copy the `profile_out_{slurm_jobid}` to your computer and visualize it locally. + +#### Metric Output +The profiled output can be analysed as follows.: + + gssr analyze -i ./profile_out + +#### PDF File Output with Plots + + gssr analyze -i ./profile_out --report + +A/Multiple PDF report(s) will be generated. + diff --git a/docs/software/gssr/index.md b/docs/software/gssr/index.md new file mode 100644 index 00000000..68aabfa4 --- /dev/null +++ b/docs/software/gssr/index.md @@ -0,0 +1,14 @@ +[](){#ref-gssr-overview} +# gssr + +GPU Saturation Scorer (gssr) provides a simple way to profile your code and get the results in both tables and plots for easy visualisation. gssr works on top of [NVIDIA Data Center GPU Manager (DCGM)](https://developer.nvidia.com/dcgm) and thus only NVIDIA GPUs are currently supported. + +The following documentations will be available.: + +* [Quickstart Guide][ref-gssr-quickstart] +* [Container Guide][ref-gssr-containers] + +This tool will produce time-series and heatmaps of the profiled metric values. Here is an example of one set of plots generated by the tool from the application Megatron-LLM from EPFL. + +![gssr timeseries](../../images/gssr/timeseries_eg.png) +![gssr heatmap](../../images/gssr/heatmap_eg.png) diff --git a/docs/software/gssr/quickstart.md b/docs/software/gssr/quickstart.md new file mode 100644 index 00000000..655eff93 --- /dev/null +++ b/docs/software/gssr/quickstart.md @@ -0,0 +1,54 @@ +[](){#ref-gssr-quickstart} +# gssr - Quickstart Guide + +## Installation + +### From PyPi + +`gssr` can be easily installed as follows.: + + pip install gssr + +### From GitHub Source + +To install directly from the source: + + pip install git+https://github.com/eth-cscs/GPU-Saturation-Scorer.git + +To install from a specific branch, e.g. the development branch, from the source: + + pip install git+https://github.com/eth-cscs/GPU-Saturation-Scorer.git@dev + +To install a specific release tag, e.g. gssr-v0.3, from the source: + + pip install git+https://github.com/eth-cscs/GPU-Saturation-Scorer.git@gssr-v0.3 + +## Profile + +### Example + +If you are submitting a batch job and the command you are executing is: + + srun python abc.py + +The corresponding srun command should be modified as follows.: + + srun gssr profile -wrap="python abc.py" + +* The `gssr` option to run is `profile` +* The `"--wrap"` flag will wrap the command that you would like to run +* The default output directory is `profile_out_{slurm_job_id}` +* A label to the output data can be set with the `-l` flag + +## Analyze + +### Metric Output +The profiled output can be analysed as follows.: + + gssr analyze -i ./profile_out + +### PDF File Output with Plots + + gssr analyze -i ./profile_out --report + +A/Multiple PDF report(s) will be generated. diff --git a/mkdocs.yml b/mkdocs.yml index bb8b560c..8a6c90e9 100644 --- a/mkdocs.yml +++ b/mkdocs.yml @@ -82,6 +82,10 @@ nav: - 'Building uenv': software/uenv/build.md - 'Deploying uenv': software/uenv/deploy.md - 'Release notes': software/uenv/release-notes.md + - 'gssr': + - software/gssr/index.md + - 'Quickstart Guide': software/gssr/quickstart.md + - 'Container Guide': software/gssr/containers.md - 'Debugging and Performance Analysis': - software/devtools/index.md - 'Using NVIDIA Nsight': software/devtools/nvidia-nsight.md