Compilation

On Taurus load the required packages (order is important!):

mkdir build
cd build
cmake .. -DCMAKE_BUILD_TYPE=Debug

To generate visualization from plots you need R. With all the packages this will require about 100 Mb of space on your disk.

In Debian you would need to install following packages:

• r-base
• r-cran-data.table
• r-cran-ggplot2
• r-cran-tidyr
• r-cran-reshape2
• r-cran-optparse

Enabling tests

By default the simulator compiles without tests. If you want these, add -DUNIT_TESTS=ON to cmake configuration.

Usage

Use "--help"

Possible parameters for --faults are "none" and "uniform". If you want explicitly specify list of failed nodes pass the node ids in comma separated list, like --faults 1,4,7.

Possible parameters for --coll are kary_bcast, checked_corrected_kary_bcast, etc. For the full list run the simulator with --coll list.

Example usage

./flogsim -P 128 -o 1 -L 4 --faults uniform -F 4 --coll checked_corrected_kary_bcast

Result printing

Please note that output formats print slightly different information. The reason for that is purely practical.

Format 'csv' prints the results in a form of csv table. List of failed nodes is not printed in this format.

Format 'csv-id' allows to save on experiment description. Instead the user is expected to provide additional token, which identifies the experiment as parameter '--id'.

Format 'csv-columns' requires additional parameter --header with a coma separated list of column names. If a column is not known, the simulator prints an empty value when the result is generated.

Visualisation

To be able to visualise the trace, add parameter --log when running the simulator. The argument sets the prefix for the trace, for example: --log log.

You can use ../script/plot_trace.R script to visualize generated traces. Run the script with --help to see all available options.

The typical usage will be:

../script/plot_trace.R -m ./log.model.csv ./log.trace.csv

Will by default create file plot.pdf out of model log.model.csv and trace log.trace.csv.

If you are overwelmed with number of messages, you can either disable them using -c flag, or only show specific nodes using --messages flag.

If you run the simulator in repeat mode (makes only sense with failure injections), trace log get additional prefix to mark run number.

Fault injection experiments

Simulating fault injection experiments (FIE) requires significant computation resources. That is why the way to do them is to run experiments on taurus.

FIE requires a long running MySQL/MariaDB server running. I do not describe the way to install the MySQL database. I just assume that you have one installed.

The whole fault injection can be done by running script fault_tests.sh.

Here is what it does.

1. Sets BASE directory
2. Starts the server (faults_server.sh)
3. Queries the server name
4. Optionally resets the database (faults_create_plan.sh)
5. Start the tests (faults_run.sh)

Look onto exported variables in fault_tests.sh to configure FIE parameters.

Be aware that the scripts were written with SLURM in mind.

TODO

• Write tests
• Add various finishing events: All done, All received, etc.
• Add ceil(L/g) network capacity limit

Dropscon

• Exponential backoff

Code style

Constructors

Write constructors with initialization lists as:

class A
{
  int c, d;
public:
  A()
    : c(0),
      d(0)
  {}
};

Mind the colon!

Whitespace

• Put space after coma