This message was deleted.

Thank you for the report @refined-telephone-69723. We have not had use cases of heavy data channel usage. Also, no large packet usage. We will have to investigate the limits. From a theoretical standpoint, losses in any of the path is going to affect throughput. Can you please elaborate on your use case? How much data do you expect to push through?

I would say our largest payload size being around 2-5MB, with regular (median) payload sizes being around 10-50kb. Given the 64kb packet limit (that seems inherited from Pion?), we chunk packets bigger than 60kb so if we sent a 3MB file, we would send ~50 60kb packets

Thank you Jimmy. Throughput wise, are you looking at an upper limit of 5 MB (approx. 40 mbps) or can the 5 MB payload span across several seconds?

When we do send those large chunked payloads, we see 2mb/s throughput in multi-region setups, which is really slow. Would like to see at least 10-20mb/s if the client can support it but I think it can be reasonably much faster (100mb/s)

jinx, just sent the question above. Thank you for elaborating on the requirements 🙂

Haha thanks for the help. At the risk of saying 1200 unbatched RPCs, I suspect there is a chokepoint with forwarding between SFUs regions / suboptimal buffering. Throughput is really important for us and one of the main reasons we choose to integrate with LiveKit, we want to be able to send data reliably and quickly (not limited by client upload speeds with many peers without SFU arch, but now it seems to be limited by LiveKit throughput)

Interesting, are you using

RoomService

to ingest data into the pipeline? Can you share the throughput you get when not using multiple region? We do not have buffering , but I am wondering if any loss is causing back pressure and a significant drop in throughput?

No ingest! Just from Livekit JS client -> another Livekit JS client. Let me test same region throughput, my memory is ~2-3x faster, I’m testing by sending 5MB of data across the network (chunked, 60kb each)

Yeah, my napkin math came out to 14😊. We have a task filed internally for this. We will discuss and investigate

Thanks!

We have not had a chance to look at it yet Jimmy. Hopefully, in the next couple of weeks.