cu_intercept

A small preload utility to intercept CUDA memory managment.

---

Compile

cmake -B build .
cd build
make

Usage

LD_PRELOAD=<path to libcuintercept.so> <your cuda programm>

Outputs

In the path where you are running the programm with the LD_PRELOAD you wil find files named output.<pid>.txt where <pid> is replace by the pid of your programm. They contain the amount of memory allocated by the programm at every time of allocations.

you can read them with python like so:

import numpy as np
import matplotlib.pyplot as plt

out = np.loadtxt("output.324547.txt")
plt.plot(out[:,0],out[:,1])

plt.xlabel("time [s]")
plt.ylabel("memory [byte]")
plt.show()