DCSim

Simulator for the simulation of high energy physics workloads on distributed computing systems with clusters, worker nodes, storages and caches.

Install instructions

You can use the Docker images available at https://hub.docker.com/repository/docker/mhorzela/dcsim/general to run your simulations. Example configuration files are embedded in the ~/data directory of the image.

If you want to build the simulator your own, checkout this repository and follow the subsequent steps.

Option 1

To get a fresh installation on your own system, either git clone this repository and execute the checkout script inside:

source checkout_scripts/checkout_with_sudo.sh

Mind that you will need super-user rights to do so, as well as cmake, git, clang and boost installed on your system. This will install the executable dc-sim for this simulator and all its software dependencies.

Option 2

With a conda environment, you would be able to install the full software setup without super-user rights. More information on how to work with and develop in a conda environment can be found in the Conda Documentation

Create a conda environment using the provided script

checkout_scripts/install_conda_environment.sh

This will automatically take care of all the dependencies needed and include them into the environment.

To work with this environment interactively, you first have to initialize conda on your system. This can be done via:

</path/to/your/conda/installation>/bin/conda init

This would adapt your ~/.bashrc to be able to call conda directly. So please re-open your shell or source ~/.bashrc.

To activate the environment, execute

conda activate dcsim-env

and deactivate it accordingly with

conda deactivate

Tips for sharing your Conda environment

It is possible to put a complete conda environment into a tarball to be able to export it to a different machine, e.g. a batch system node. To do that execute:

conda activate dcsim-env # in case you don't have it activated yet
conda-pack

The created tarball dcsim-env.tar.gz can then be uploaded to a storage element and copied from there to a different machine.

Beware: It was observed that the simulation from a shared Conda environment can be significantly slower than in the native installation.

Usage

When you have successfully installed the simulator or activated the conda environment you can run

dc-sim --help

to see all possible execution options.

Mandatory parameters are a platform file and a path and name for the resulting simulation output CSV-file:

dc-sim -p <platform-file> -o <output-path> --workload-configurations <path_to_workload_json> --dataset-configurations <path_to_dataset_json>

The platform file has to follow the SimGrid-defined DTD. Example files can be found in data/platform-files. The output-path can be any relative or absolute path of your file-system where you are allowed to write to. Instead of manually setting up all workload and dataset parameters, JSON files can be provided, which contain all necessary information

An example for a workload mixing both gaussian and histogram distributions for its job characteristics would be, e.g.:

{
    "calc_workload": {
        "num_jobs": 60,
        "flops": {
            "type": "histogram",
            "bins": [1164428000000,2164428000000,3164428000000],
            "counts": [50,50]
        },
        "memory": {
            "type": "gaussian",
            "average": 2000000000,
            "sigma": 200000000
        },
        "outfilesize": {
            "type": "gaussian",
            "average": 18000000,
            "sigma": 1800000
        },
        "workload_type": "calculation",
        "submission_time": 0,
        "infile_dataset": "calc_dataset"
    }
}

It is also possible to give a list of workload configuration files and configure more than one workload per file, which enables to simulate the execution of multiple sets of workloads in the same simulation run. Example configurations covering different workload-types are given in data/workload-configs/.

The dataset configuration file must contain locations of the data, number of files in the dataset and file sizes, defined via a probability distribution. An example for a dataset mixing would be, e.g.:

{
  "calc_dataset": {
      "location":["RemoteStorage"],
      "num_files": 0,
      "filesize": {
        "type": "gaussian",
        "average": 0,
        "sigma": 0
    }
  },
  "second_dataset": {
      "location":["FirstHost", "SecondHost"],
      "num_files": 100,
      "filesize": {
        "type": "histogram",
        "counts": [210319, 577, 0],
        "bins": [0, 20, 40, 60]
    }
  }
}

Example configurations covering different dataset-types are given in data/dataset-configs/.