This repository contains all material used to perform the experiments presented in the paper "Affinity-Aware Resource Provisioning for Long-Running Applications in Shared Clusters".

Input instances generation

The Python scripts to format and generate instances from the TClab dataset are in data/scripts. The scripts needs the 3 following CSV files from the TClab dataset (available in data/TClab:

• app_resources.csv
• app_interference.csv
• instance_deployment.csv

The datasets TClab_dataset_2D.csv and TClab_dataset_TS.csv are generated from these 3 files using the script generate_TClab_dataset.py. These datasets serve as baselines to generate all instances of higher density and large scale, both for fixed and time-varying resource requirements, using the scripts generate_higher_density.py and generate_large_scale.py.

Building executables

The only dependency for building executables is cmake. We use Nix as a package manager. default.nix file contains a recipe to build all executables needed. With Nix installed, simply run nix-build default.nix -A binpack and nix will do the rest and create a result/bin folder with all built executables.

If you want to be in an environment to hack the C++ algorithms and manually build the executables, run nix-shell default.nix -A binpack instead.

Input file format

The input datasets are TAB-separated CSV files containing 6 columns. Apart from the header line, each line in a file describes an application. The 6 columns are:

• app_id: The unique identifier of the application
• nb_instances: The number of replicas of the application
• core: The core requirement of the application
• memory: The memory requirement of the application
• inter_degree: The number of affinity restrictions of the application
• inter_aff: The list of affinity restrictions, formated as a string representation of a Python list. The list must have inter_degree elements, and each element is a couple (application_id, affinity_value) (using notations of the paper, this corresponds to the couples $(j, a_{ij})$ for application $i$)

Output file format

Each executable generates an output file in data/results, which is a TAB-separated CSV file. Apart from the header line, each line in a file records the results obtained by each algorithm applied to a given dataset instance. The columns are:

• instance_name: The name of the dataset instance
• LB: The value of the lower bound
• best_sol and best_algo: The best solution value found, and the algorithm name that found it. If multiple algorithms found the best solution, only the first algorithm is kept.
• One column per algorithm name, with the solution value found by this algorithm
• One column per algorithm name and suffixed with _time, with the time (in seconds) taken by the algorithm to find the solution