Bankers-Algorithm-CUDA

CUDA Implementation of Banker's Algorithm for Deadlock Avoidance

Please read the following instructions carefully in order to execute the code. See Technical Report for technical details.

Requirements:

• CUDA 5.0 or higher
• NVIDIA GPU with compute capability 3.5 or higher
• Windows environment (due to benchmarking functions using <windows.h>)

Visual Studio Settings

For the GPU implementation we utilize Dynamic Parallelism. Programs with Dynamic Parallelism require additional project settings in the Visual Studio in order to be compiled and run:

1. Right click project in the solution explorer and go to Properties.
2. Under Configuration Properties -> CUDA C/C++ -> Common set Generate Relocatable Device Code to Yes (-rdc=true)
3. Under Configuration Properties -> CUDA C/C++ -> Device set Code Generation to compute_35,sm_35
4. Under Configuration Properties -> Linker -> Input -> Additional Dependencies add cudadevrt.lib
5. Make sure that execution settings (Debug/Release option) is the same as the configuration in the Properties window.

Source: https://stackoverflow.com/a/59383269/9817067

Notes about execution of the program

• Program requires allocation.txt, available.txt, max.txt, need.txt, request.txt, and info.txt to be available in the directory of the executable. Modify the paths under readMatrices() function accordingly if you would like to store datasets in somewhere else.

• Datasets used during experiments are provided in the DATASETS.rar file. You can unzip that file, pick a dataset and copy 6 text files to the execution folder in order to test them. My experimental results of each implementation in all datasets are also provided in results folder of each dataset.

• Note that depending on the dataset size, loading matrices from the disk to the memory may take a long time. Please be aware.

• There is a flag named multiThreaded in the beginning of the main function of CPU implementation. Set this flag to true if you would like to execute partially-parallel version of CPU implementation. When it is set to false, all operations are performed sequentially, in a single thread.

• Algorithm outputs whether the requested allocation is servable or not by printing a message to the terminal. In order to print some intermediate data structures you can uncomment necessary lines in the codes or add your own print statements.