Hacker News Reader

Comment by mikeash | Hacker News Reader

Since we're all sharing wacky debugging stories....

I had a user of an audio streaming product who suffered from random disconnects. He could only stream for about five minutes before it quit. But, and this was the crazy part, the problem only happened when streaming rock and roll. Classical music worked fine.

After many weeks going back and forth, I tracked it down to a misconfiguration on his network interface. The MTU was set lower than the IPv6 spec allowed (or so, forgive me if I get this detail wrong) and the OS wouldn't fragment our UDP packets with this setting.

Our packets usually were under the limit, but occasionally went over and triggered an error. Why was classical music immune? The lossless codec we were using did a better job on classical than rock, so classical (or at least this guy's classical) never went over the size limit.

View on HN

I'm assuming this is due to the lower average signal level of symphonic recordings? Rock music is mastered closer to 0db due to the well-publicized loudness war.

Obviously, this doesn't affect bandwidth in uncompressed audio, where every sample contains every bit, so filesize = bit-depth * sample-rate(hz) * T(seconds). But in compressed audio, average signal level can make a significant difference in filesize/bandwidth.

(Who would have thought a network would be a casualty of the loudness war... harhar)

I think rock is just less regular at a small scale. Consider a flute versus an electric guitar. The flute is at least somewhat like a pure tone which would be easy to compress, while an electric guitar is a mishmash of stuff that will look much more random on a sample by sample basis.

Cool -- I didn't even consider the shape of the waveforms (sine-like for flute, saw/square-like for anything distorted) having an effect on the data, but now that you tell me... duh!

To be clear, I surmised this from what I described, I didn't actually get down and poke at how the codec works on these different types of music. But I'm pretty sure that's what's going on.

These lossless codecs work by attempting to predict the next sample from all the previous ones, and then encoding the difference between the prediction and reality. So a more regular waveform will compress better than a crazy-looking one. It's somewhat counterintuitive since rock music tends to be less complicated from a high-level human perspective.

Rock music has more information (in Shannon's sense) than classical music.

Yeah, I imagine all those distortion and amp units introduce a lot of fairly loud random noise.

Reply on news.ycombinator.com