GPU Project Predicting the Performance of a kernel

Setups

1. SSH TO A CUDA SERVER
   • Example (using GPU 2):

     ssh cuda2.cims.nyu.edu
     

2. LOAD CUDA MODULE
   • Example:

     module avail cuda
     module load cuda-12.2
     

CUDA Benchmark Program — GPU_project.cu

Compile:

nvcc -o project GPU_project.cu -lm

Run commands:

./project [PROBLEM_SIZE] [NUM_THREADS] [NUM_BLOCKS] [OPERATION]

For example:

./project 100 32 5 addition

Available Operations:

addtion, subtraction, multiplication, and reduction

It should output something like

Number of devices: 2
Currently using Device: 0

Device 0:
name: NVIDIA GeForce RTX 2080 Ti
total global memory(KB): 11081664
shared mem per block: 49152
warp size: 32
clock rate(KHz): 1635000

Single thread time = 5.041408 secs
GPU: 5 blocks of 32 threads each
Kernel time = 0.039840 secs
Speedup: 126.54x

Automation Script — experiment_automation.py

The automation script tests CUDA kernel performance across different configurations and GPUs. It repeatedly runs the compiled CUDA program (gpu_project) with varying parameters (problem size, threads per block, number of blocks, and operation type), records the output, and writes averaged timing results into a CSV file.

What It Does

• Executes the CUDA binary using different configurations.
• Captures printed output (CPU time, GPU kernel time, and computed speedup).
• Parses those values using regular expressions.
• Repeats each experiment multiple times to average out timing noise.
• Saves all averaged results to a file named results_.csv (e.g. results_cuda2.csv).

Each GPU node generates its own CSV file, which can later be merged into one master dataset for ML analysis.

How to run the automation script:

RUN THE AUTOMATION SCRIPT:

• replace the (insert_name_of_operation) with either addition, subtraction, multiplication, or reduction

python3 experiment_automation_(insert_name_of_operation).py