Files

• mc321.c: Base implementation in C.
• mc321.chpl: The experimental Chapel version that can run on GPUs.
• mc321.out: 10M photon output of the base implementation
• mc321.out.chpl: 10M photon output of the first Chapel port that's not parallel
• mc321.out.chpl.gpu: 10M photon output of the GPU-enabled Chapel version

How to compile

> gcc -lm -O3 mc321.c   # for the C version
> chpl --fast mc321.chpl -suseCudaRng=true  # for the Chapel version

Notes

• The Chapel version wants to use the custom RNG by default. However, there are some issues launching the kernel with that. Instead, -suseCudaRng=true can be used to use CUDA's RNG, which works fine.

Performance

Below are some results. The Chapel version uses 1M GPU threads. All runs are using 1B photons.

RTX A2000 is a low-end GPU, whereas A100 is an HPC-grade GPU. There are more powerful ones that I can test with as well.

	Sequential C	Chapel w/ RTX A2000	Chapel w/ A100
Time (s)	83.57	6.67	0.80
Throughput (MPhotons/s)	11.97	149.94	1254.21
Relative Performance	1.0x	12.53x	104.78x

• Chapel 2.2 (compiled with only --fast)
• gcc 13.2 (compiled with -O3 -lm)

Pushing the limits

On A100, I wanted to see how high the throughput can get and played around a bit with number of photons and threads. With 100B photons and 100M threads, I get the following, which was the highest throughput I've seen so far:

Number of photons : 100000000000
MPhotons/s : 1384.64
Elapsed time : 72.2208

This is about 115x better throughput than the sequential C version.