General

To get access to the RCI cluster: https://docs.google.com/forms/d/e/1FAIpQLSewws_V6-D567fkp6QZmr0GQlkzQrEoB6QquAgQkZu8so818Q/viewform

Official instructions: https://login.rci.cvut.cz/wiki/how_to_start

Usual command Usage

  • You can watch your jobs with squeue -u username.
  • You test/debug your program when running it with srun command
    • you usually don’t have to allocate resources when testing
    • To start an interactive shell, run: srun -p cpufast --pty bash -i
    • Set your main script file (sh) executable via chmod +x <filename> command
    • Test your script in console
    • Cancel the job: scancel <JOB ID>
  • You run your job with sbatch command with allocated resources
    • Example sbatch: sbatch --mem=30G -t 60 -n30 -o /home/fiedlda1/Amodsim/log/ih.log /home/fiedlda1/Amodsim/rci_launchers/ih.sh

How to clone projects

Usually, you need to clone some of the SUM projects to start working on the RCI cluster. To do that:

  1. Copy your key to ~/.ssh/
  2. Set file permissions to your ssh key safely
  3. Modify ~/.ssh/config IdentityFile to point to your key
  4. Clone your project

Specifics for java projects

  1. Clone project on RCI cluster
  2. Download binary maven from: http://maven.apache.org/download.cgi and export it to your home folder on the RCI cluster.
  3. Prepare your bash script
    • Add #!/bin/bash to the first line
    • Change the environment variable PATH for your maven location with this command: PATH=$PATH:/home/$USER/apache-maven-3.6.1/bin/
    • Load all required software via ml command, definitely ml Java
    • Build, compile, run your project via mvn commands
    • Set your file executable via chmod +x filename command
    • Example run command: mvn exec:exec -Dexec.executable=java -Dexec.args=’-classpath %classpath -Xmx30g cz.cvut.fel.aic.amodsim.OnDemandVehiclesSimulation /home/kholkolg/amod-to-agentpolis/local_config_files/olga_vga_RCI.cfg’ -Dfile.encoding=UTF-8
    • Example - Bash script for amodsim project:
  4. Run your script with srun, sbatch etc. commands, I recommend first use srun, to check everything is set up ok and then use sbatch command, because if computational nodes are busy, your job will be added to the queue and you can do other work.

Specifics for python projects

  1. First load the appropriate version of python, e. G.: ml Python/3.6.6-foss-2018b
  2. You can’t just install the packages with sudo, you have to install them to the user space instead:
    • Run pip install --user packagename

Specifics for C++ projects

Workflow options

As linux binaries are usually not portable. They are not compatible with older linux versions due to the infamous glibc incompatibility. There are three solutions to this problem:

Method Setup Program Upgrade
Compile the code on the RCI Setup the compilation on RCI. Copy the source code to RCI and recompile after every change
Use a Singularity container learn with singularity, create the container Generate new container and copy it to the RCI
Build a compatible binary using a modified toolchain learn with a toolchain generator, configure and generate the right toolchain Copy the updated binary

Building on RCI

In general the workflow is the same as on a local machine. The difference is that we do not have root access, so for all needed tools, we have to either load them via ml command, or, if not available, install them in the user space. Typically, we need to load:

  • git: ml git
  • GCC: ml GCC
  • CMake: ml CMake

Specific for projects with gurobi

  1. Load Gurobi with
    • ml Gurobi
    • or ml Gurobi/8.1.1-foss-2018b-Python-3.6.6 for a specific version
  2. Be aware that this operation can reload other packages

Gurobi and Java

It is necessary to install Gurobi to maven: mvn install:install-file -Dfile=/mnt/appl/software/Gurobi/9.0.3-GCCcore-8.3.0-Python-3.7.4/lib/gurobi.jar -DgroupId=com.gurobi -DartifactId=gurobi -Dversion=1.0 -Dpackaging=jar

Gurobi and C++

As RCI use Linux as OS, we need to compile the Gurobi C++ libs with the same compiler as the one we use for compilation of our code (see C++ Workflow for more details). Note that this is necessary even if the Gurobi seems to be compiled with the same copiler we use for compilation.

Unlike in Linux installation we controll, we cannot build the C++ lib in the Gurobi installation folder. To make the Linking work, foloow these steps:

  1. copy the src dir from the RCI Gurobi module located at mnt/appl/software/Gurobi/<desired version> to our home
  2. run make located in src/build
  3. copy the libgurobi_c++.a to the lib subfolder of your project

  4. configure the searching for the C++ lib in FindGUROBI.cmake file:

    ```cmake

    Find the Gurobi C library

    find_library(GUROBI_LIBRARY NAMES gurobi HINTS ${GUROBI_DIR} $ENV{GUROBI_HOME} PATH_SUFFIXES lib )

    Find the Gurobi C++ library

    find_library(GUROBI_CXX_LIBRARY NAMES gurobi_c++ HINTS ${PROJECT_SOURCE_DIR} PATH_SUFFIXES lib NO_CMAKE_ENVIRONMENT_PATH REQUIRED )

    Find the Gurobi C++ debug library. We don't need it on RCI, but it may be required in CMakelists.txt. Therefore, we supply the release version as a debug version.

    set(GUROBI_CXX_DEBUG_LIBRARY ${GUROBI_CXX_LIBRARY})

    ```

  5. if the CMake cache is already generated, delete it.

  6. Generate the CMake cache and build the project

Commands

For command description, see Slurm manual.

Main params

  • --mem=<required memory>
    • <required memory> is in megabytes by default, for gigabytes, we need to add G
  • -t <time>
    • Time in minutes

Task Explanations

srun --pty bash -i - --pty runs the first task and close output and error stream for everything except the first task - bash: what we want to run - -i Input setting, here followed by no param indicating that the input stream is closed osed

Access the RCI cluster using JetBrains Gateway

When connecting to the RCI cluster, we have to use the login2 or login3 nodes. The login1 node has an outdated version of glibc which is not compatible with the JetBrains Gateway.

Resource limits

The resource limits are described in the RCI cluster documentation. However, the limits:

  • do not cover the amd partitions, and
  • there is no information how to determine the role for a user.

Therefore, to know exactly the limits, it is best to use the sacctmgr show command.

Installing Linux packages

Without root access, we have cannot install packages using the package manager. However, we can manually install them to home. Steps:

  1. Locate the repository of the package that contains the packages for Red Hat Linux (rpm)
  2. Download the package
  3. Extract the package using rpm2cpio pv-*.rpm | cpio -idmv
  4. Move the extracted files to the desired location in the home folder
  5. Add the bin folder of the package to the PATH environment variable