FlavioK / cuda-launchoverhead

Short experiment to measure the launchoverhead of CUDA kernels on the Jetson-TX2

Launch overhead measurements

This is a small experiment to measure the launch overhead of CUDA kernels on the Jetson-TX2. Mulitple spinning kernels are launched from different streams. To measure the overhead the time before the launch is recorded on the CPU. After the kernels starts executing the timestamp of the globaltimer is taken and converted to CPU time. The difference of these start times represent the launch overhead.

Configuration

Set device number, used number of streams (realized with cudaStreamPerThread) and number of kernel repetitions per stream.

#define DEVICE_NUMBER (0)
#define NOF_STREAMS (6)
#define NOF_REP (3)

Set spin duration of a single kernel

#if 1
#define SPIN_DURATION (30000) // 30us
#else
#define SPIN_DURATION (30000000) // 30ms
#endif

Synchronize with the stream after each kernel launch.

#if 1
#define STREAM_SYNC() CheckCUDAError(cudaStreamSynchronize(cudaStreamPerThread));
#else
#define STREAM_SYNC()
#endif

Building

If you clone directly to the device call:

make all

and then run

./launch-overhead

If you have ssh access to a Jetson-TX2 with an existing nvidia account call:

make target_run HOST=tx2fk

This will copy the project to the device via ssh and compile and run it on the target.