Memory Benchmarks for Cupy

Test CPU - GPU copies vs cupy's unified memory

Python Files

• implementations.py : contains gpu_copy and unified   functions

  • gpu_copy : memory copies between CPU - GPU
  • unified : use of cuda unified memory ( by using unified_{pointers, arrays} functions )
  
  

• timers.py : contains helper function to time given function and returt result and stats

• test.py : manually test function from implementations.py with below parameters
  • size: Array size (NxN) used for testing (default=10)
  • repeat: Benchmark repetition times (default=10)
  • num_oper: Perform #num_oper operation loops (default=10)
  • gpu_copy: use of gpu_copy function
  • unified: use of unified function

• bench.py : Bench the above-mentioned functions and plot results

Results

GPU specs used for experiment with bench.py:

nvidia-smi

Benchmark for different array sizes

Benchmark for different number of operations loops