Following a discussion with Mark, I made some TCP settings changes to
hemlock to use the Wikimedia standard linux network stack tunings. I
do not expect any negative impact from this.
I also enabled TCP ECN
(
http://en.wikipedia.org/wiki/Explicit_Congestion_Notification), which
is now also enabled on a number of other non-critical wikimedia
services. Just like our use of IPv6 can cause problems for broken
clients, the same risk exists for TCP ECN. A number of fairly high
profile sites (such as
kernel.org) have run TCP ECN, so I also do not
expect any problems from this change.
If you find anyone who is suddenly unable to connect to toolserver and
is also unable to connect to
http://vger.kernel.org/ecn-on.html, then
they may need to fix their firewall. I'd love to hear about any such
cases. Thanks.