thank you so much for you hints, I’ll definitelly give them a try.
yes, kind of, they started like a garage company (part of original comunity network CZFree Net ) and built their infrastructure on mikrotik routers
I defintely had to turn it off, otherwise they won’t unban me And I turned it on just recently because the ISP guys are off-duty during the night so I may experiment a bit without being caught
btw. I captured a package dump of about 1000 sec of my (mostly IPFS) network trafic (if you are interested): https://gateway.pinata.cloud/ipfs/QmVjHXJVvGwzEG9YZb4ZmPDzjVYgE3HgJ7zMaBJ5wy6Y9T
and this was quite interesting (if I understand the dump correctly):
- i kept
Swarm.ConnMgr.HighWater to 200 but the number of peers was all the time at arround 400
(so the ConMgr had obviously hard time to keep-up with the incomming connections - WHY ?)
- once I started go-ipfs the packet loss went from 0 to ~30% at the IPS’s internet gateway
(so likely a trafic shaping)
- 64 % of all packets had size between 80-159 bytes (so mostly DHT trafic it seems)
- only 2% of all packets were “large packets” (so I likely am not a relay)
- multicasts/broadcasts were only 0.6% of all packets and MDNS only 0.03% (so the teory that the wifi was clogged by slow multicast frames is wrong)
- the IPFS daemon actually talked with ~ 4000 IP addresses over UDP where about 2/3 of these connections only exchanged 3 packets ! and with roughtly 7000 IP adresses over TCP where the median is ~119 packets per connection (43 packets send and 77 received).
- 37% of all packets were actually TCP failures, like duplicate ACK’s, re-transmited frames etc.
so my new theory is that the root-cause of this problem is actually a trafic shaping at ISP’s side, which is seriously disrupting those miriads of short-lived connections, which then respond with re-sending the trafic, which leads to packet multiplication and eventual congestion of the wireless backbone link.
Any cures for that ? E.g. any way how to make my ipfs daemon less desirable peer for others ?
I’m going to test it with DHT in client-only mode as recommended by lidel and see how it goes.