[tor-relays] numtcpsock prevents relay from working properly.

["Nils Erik Flick" <Flick@xxxxxxxxxxxxxxxxxxxxxxxxxxx>]

Hi,

I am experiencing trouble with a new VPS relay.

 https://atlas.torproject.org/#details/41F07731207742860D43AC426FBAE2F3947BD1CA

It spontaneously acquired an advertised bandwidth of 2.67MB/s but then
started to sputter.

 Failing because we have 1314 connections already ...

Specifically, looking at

 /proc/user_beancounters

numtcpsock is the culprit, which is limited to 1360 (some other
processes are using the remaining sockets). This artificial limit causes
new connections to fail even though send/receive buffers or CPU usage
are well below critical.

The value can't be changed from inside the OpenVZ container. That's a
pity because traffic appears to be unmetered and a lot of potential is
wasted this way.

I understand that there is currently no way to limit this inside the Tor
process because every relay needs to be ready to connect to every other
relay at any time.

So I tried to get its consensus weight down. Gradually lowering the
MaxAdvertisedBandwidth to 300k caused it to briefly acquire the "Stable"
flag and run into trouble again.

That's no better than a home connection. Worse, in fact, because the
problem persists.

I'm curious, is this a common experience?

Asking customer service would mean explicitly drawing attention to the
fact that I'm trying to use the cheap plan to run Tor nodes. That's
completely against my instincts. On the other hand they *did* offer a
free-form box in the registration form.

Has anyone tried and if so, to any avail?

Is there a workaround I'm missing? iptables can be configured freely,
but I don't see how that would help.

At this point I'll probably call it a day, try elsewhere and set this
one up as a bridge instead.

Frankly I was a bit surprised that technical snags like this aren't
generally mentioned in relay setup guides or FAQs. For all I know, it
may be common knowledge among expert sysadmins that OpenVZ should be
shunned or whatever, but to someone with little prior experience as a
user of commercial hosting, some of these limitations come as a surprise.

In a bout of post factum searching, I only found this discussion

https://serverfault.com/questions/698845/debian-limit-number-of-tcp-sockets-for-a-process-especially-tor-on-openvz

IMHO this kind of knowledge should be spread more "pro-actively", which
is why I'm posting this on a public list.

Peace,
Erik
_______________________________________________
tor-relays mailing list
tor-relays@xxxxxxxxxxxxxxxxxxxx
https://lists.torproject.org/cgi-bin/mailman/listinfo/tor-relays