Debugging transfer speed

[vanilla-plus]

Is there any way to debug connection speed issues?

I just tried out sendme with a friend; my connection is 198Mbps down, 42 up. Theirs is 75 down, 17 up.

Me sending them a file peaked around ~500Kbps. Them sending me a file peaked at 1Mbps.

We're in the same city and there doesn't seem to be many options to configure sendme, but the poster site said it could saturate 4Gbps.

Are there any common factors that could impact transfer speeds?

[matheus23]

It's fairly hard to debug these kinds of issues.
One very basic thing to try is to see what kind of connection you end up having (i.e. via relay, or direct). To test this, you can use iroh-doctor: https://github.com/n0-computer/iroh-doctor/
But you'd need to compile from source.

Then you set up one side with cargo run --release -- accept --iterations 1, and the other side will need to run cargo run --release -- connect <node_id>

This will print when the "Connection type" changes - i.e. from Mixed (relay + direct) to Direct, and after what time.
It'll also measure transfer speed.

If it turns out you're connected directly, then the issue can come from many places:

• ISPs may route traffic inefficiently for reasons out of control for iroh
• Although iroh can in principle saturate 4Gbps (and we've done so in the past), not all code that uses iroh will do that. E.g. you can improve transfer performance by implementing reading data chunks out-of-order or in parallel with multiple buffers. These optimizations aren't implemented in sendme/iroh-blobs yet.

But alas - the first thing to check is if traffic is routed directly or relayed.

[ambyjkl]

This software struggles to maintain 3MiB/s between two 10G servers in US east coast and Europe and barely manages 40MiB/s between two severs in the same location with sub 1ms ping. I did manage to "replicate" your 4Gbps claim tho using a top-secret setup, sending from from localhost to localhost... yeah. This software needs some work.

In the localhost to localhost test, it uses about 20% CPU split between multiple cores for some reason. I'm not sure if it's just blake3 checksums because b3sum takes about 1/50th the time to operate on the same 1GB file.

If you want to optimize performance, I'd recommend looking into tokio to begin with, since the 20% load spread out over all cores screams "work stealing". I think just running in a single thread would be faster tbh.