Processing Speeds

[djohorton]

Hello again,

I've followed the Reddit thread regarding how to increase processing speeds for the BirdNET GUI, prompted by a discussion I created on the BirdNET Analyzer github page. I went ahead and got access to high performance computing here on my university campus, learned some basic Linux coding, and got BirdNET-Go running.

Like I said in my previous discussion about processing speeds -- a problem, as I have at least 8 TB of data -- I was only able to run a batch analysis of ~10 GB of data at a time, taking about 2.5 hours for the batch run in the GUI. On BirdNET-Go, it's taking the same amount of time for the same amount of data. In the read.me for BirdNET-Go, I read that that time it takes to "process a 3-second audio chunk... should range from 50ms to 550ms, depending on the hardware you are running BirdNET-Go on. As long as the processing time is less than 1500ms, audio is processed quickly enough to keep up with real-time capture and analysis."

For reference, these are the settings I use when I begin an interactive shell session:

• normal partition
• walltime of 24 hours
• 32 total CPU threads
• 32 GB of RAM

I am positive there is something I am missing at this point. I feel like I'm doing something incorrectly, because it is taking a minimum of 5 seconds to analyze a file (15-20 chunks/sec), with another 5 second gap between all files processed. Sometimes, this becomes much longer (10-20 seconds) to process one 5-minute file -- all files are 5 minutes, so roughly 1000 files per batch analysis.

I've attached a screenshot to show my lines of code. I am open to any suggestions! I don't know a whole lot of code at the moment, but I'm learning quickly! Many thanks!

[tphakala]

That ~22 chunks/sec is pretty poor performance, my desktop PC and do ~67 chunks/sec and my macbook pro with M4 Pro CPU does around 70 chunks per second which is fastest system I have tested sofar.

What I have seen tflite model used in BirdNET-Go seems to scale poorly with high number of CPUs, so instead of running BirdNET-Go with high number of CPUs like 28 in your case, try limiting CPUs to use to lower value, test with values like 6 or 8. Adding option -j 8 sets CPUs to use to 8.

My recommendation would be to use multiple smaller virtual machines instead of single large one, like those 32 CPU and 32 GB resources would be more efficiently used if divided for 4 VMs with each 8 CPUs and say 4 GB RAM (BirdNET-Go does not use much RAM at all). Directory watch mode allows those multiple instances to monitor for audio files in a network share and they will process existing or new files automatically, more VMs will process large number of files faster.

In the read.me for BirdNET-Go, I read that that time it takes to "process a 3-second audio chunk... should range from 50ms to 550ms, depending on the hardware you are running BirdNET-Go on. As long as the processing time is less than 1500ms, audio is processed quickly enough to keep up with real-time capture and analysis."

These performance figures refer to realtime analysis mode which most people use BirdNET-Go for, in that mode BirdNET-Go listens for soundcard capture and reports detected species to a web page like this https://mokki.pobox.fi/.