[tor-relays] increase txqueuelen: reasonable or not?

[Steve Snyder]

The config page at torservers.net advises increasing the TX Queue Length from the default 1,000 to 20,000.  For a differing opinion see the assertion that this increases latencies and hampers network congestion recovery ("bufferbloat"):

http://www.cyberciti.biz/faq/gentoo-centos-rhel-debian-fedora-increasing-txqueuelen/

I'm running an exit node on a 100Mbps network (multi-core CPU on RHEL5), and typically have 600 - 800 connections at any given time.  TorStatus usually reports my throughput at 700KB to 800KB per second.

I suspect that the 20,000 recommendation pertains to Gbit Ethernet.  Still, though, is the default of 1,000 adequate for my Tor traffic?

More generally, I haven't tweaked my networking (system network buffer sizes, etc.) because it is unclear to me what recommendations are appropriate for exit nodes running contemporary versions (v0.2.2.30+) of Tor.  Reliable benchmarking can't be done because the system load varies with traffic at any given time.  Is there a way to prove empirically that this or that adjustment to the system configuration really has an effect on the performance of a Tor relay node?

Thanks.