Bad performance when using only one socket

[cazala]

Bad performance when using only one socket

[cazala]

I see, so 1 tcp connection per websocket kills the pool, and 1 tcp connection per address doesn't work at scale.. I dunno if should I optimize somehow the single tcp connection or have a configurable pool of tcp connections that get reused (say, no more than 100 websockets sharing the same tcp connection).. ideas?

[justnx]

How about TCP fastopen Support?

[korzhyk]

Looks great to share 1 socket connection for limited users. Also we need to keep in mind that pool DIFF is matter and when DIFF is low (< 2000) there are too many messages will be sent to pool and when DIFF is large there are no heavy load. So maybe a better way to implement a rate-limiter for socket connection? When messages to pool over 200-400 (depending on latency) per second we can create a new socket instance and push new messages to new socket.

[korzhyk]

And I prefer for easy solution of limit miners per socket.
For example: 3000 ws connections can be handled by 60 tcp sockets (50 ws/tcp) and there are 98% profit of saved tcp connections.

[korzhyk]

@justnx TFO is not about of "speedup" data transmission this is for fast TCP connection (handshake).

[cazala]

okay, when I have some time I'll try to refactor the code to at least initially have a configurable maximum amount of miners per socket. If this solves the problem then I can implement some logic to figure out this number automatically (ie, messages per second as you suggested)

[justnx]

yes but Perfect for proxies saving you latency and connection sockets

[procloud]

i did notice that xmrig proxy claims 100k miners at ~350 connects to pool, which is about 285 miners per pool connect. you may be able to look through that projects code to see how they are distributing this algo.

[sunk818]

Does it make sense to allocate TCP sockets for specific roles? One TCP socket for authorization & hash submission. Another TCP socket for new difficulty and new block announcements? If either socket exceeds the messages/second threshold, then create another socket...

[cazala]

This can be configured now with maxMinersPerConnection (programatically) or --max-miners-per-connection (cli)