Encryption update

[dank074]

Adds support for both aead_aes256_gcm_rtpsize and aead_xchacha20_poly1305_rtpsize

Closes https://github.com/dank074/Discord-video-stream/issues/99

[longnguyen2004]

I'm wondering about whether to include AES-256, since most sources I found seem to indicate that AES-256 is slower than XChaCha20 (unless AES-NI is used, but I don't know if Node.js even utilizes that). Will do some real world tests on my 14 years old CPU to see if there's a real impact.

For context, my current setup can easily handle streams up to 10Mbps, which is Discord's limit.

[dank074]

I'm wondering about whether to include AES-256, since most sources I found seem to indicate that AES-256 is slower than XChaCha20 (unless AES-NI is used, but I don't know if Node.js even utilizes that). Will do some real world tests on my 14 years old CPU to see if there's a real impact.

For context, my current setup can easily handle streams up to 10Mbps, which is Discord's limit.

Cool I'll wait for your results. We can definitely remove AES-255 support if it's slower since according to the Discord docs it might not always be available on their end anyway. Not sure why it's their preferred option

[dank074]

We are about ~1 month from our current encryption being deprecated, so I guess we gotta merge this soon. Have you had the chance to do some performance tests yet @longnguyen2004

[longnguyen2004]

Here are some benchmark results with openssl speed

• AMD Ryzen 5 5600H, Windows 11 23H2

Block Size	2B	31B	136B	1KiB	8KiB	16KiB
AES-256-GCM (without AES-NI)	4.76MiB/s	50.38MiB/s	169.23MiB/s	399.57MiB/s	460.91MiB/s	463.67MiB/s
ChaCha20-Poly1305	4.95MiB/s (1.04x)	66.28MiB/s (1.32x)	205.59MiB/s (1.21x)	1.18GiB/s (3.01x)	2.01GiB/s (4.47x)	2.12GiB/s (4.69x)
AES-256-GCM (with AES-NI)	12.03MiB/s (2.53x)	130.21MiB/s (2.58x)	513.4MiB/s (3.03x)	2.11GiB/s (5.41x)	3.57GiB/s (7.93x)	3.79GiB/s (8.37x)

• Intel Core i3-330M, Ubuntu 24.10

Block Size	2B	31B	136B	1KiB	8KiB	16KiB
AES-256-GCM (without AES-NI)	1.08MiB/s	13.47MiB/s	38.76MiB/s	76.14MiB/s	85.4MiB/s	85.45MiB/s
ChaCha20-Poly1305	1.56MiB/s (1.44x)	23.46MiB/s (1.74x)	71.77MiB/s (1.85x)	236.51MiB/s (3.11x)	297.14MiB/s (3.48x)	302.41MiB/s (3.54x)

Commands used for benchmarking:

$ OPENSSL_ia32cap="~0x200000000000000" openssl speed -elapsed -aead -evp aes-256-gcm
$ openssl speed -elapsed -aead -evp chacha20-poly1305
$ openssl speed -elapsed -aead -evp aes-256-gcm

Overall, ChaCha20-Poly1305 is faster than non-accelerated AES-256-GCM. I also saw this with my crude real-world test using a 10Mbps stream on the i3-330M. ChaCha20-Poly1305 has a pretty consistent 2-3% CPU usage, while AES-256-GCM is 3-4.5%, with occasional spikes of 6-11%.

As expected, AES-NI blows everything out of the water, and would be the ideal solution for supported CPUs, which is the majority of CPU produced within the last decade. For those that don't, the impact is also not as much as I thought, and is perfectly adequate for a single stream, so it's probably fine to leave AES in, and add an option to force ChaCha20 somewhere.

[dank074]

Sounds good, I'll add a config option