I am going to be MFCing the transmit timer min/slop stuff soon (because
the vast majority of complaints by users related to this issue is
on -stable).
I believe the basic concept and code is reasonable and the only real
issue is whether to make the default slop 1000ms or 200ms. I would very
much like to change the default to 200ms in -stable but I will be happy
to MFC the code to -stable with a 1000ms default to begin with, to get
it out of the way, and then persue changing the default (on stable)
to 200ms as a separate issue. The default should remain 200ms in
-current.
I've considered how best to test this change in a real environment. The
only thing I can come up with is to test it on my main machine,
apollo.backplane.com, which serves around 800 pages a day, plus handles
a good deal of email, and use netstat -s to collect statistics with
the default at 200ms and the default at 1000ms.
In regards to people who have objections to the change perhaps once it
is in stable we can get a bunch of developers, including those with
objections, to test real-life production systems for a week alternating
the slop once a day 200 ms, 1000 ms, 200 ms, 1000 ms, and collecting
netstat -s statistics as a means of determining whether a 200ms default
is reasonable. I added an additional network statistic 'unnecessary
retransmits' to netstat -s in -current and will be MFCing that to -stable
as well, which should help in any evaluation. I am personally quite
sure that no major gotchas will occur, but I am willing to allow time
for my and other people's testing to confirm my suspicion. That said,
once the statistics are collected and found to show no lasting harm
the onus would be on the people with objections to prove it otherwise.
-Matt
Matthew Dillon
<[EMAIL PROTECTED]>
To Unsubscribe: send mail to [EMAIL PROTECTED]
with "unsubscribe freebsd-hackers" in the body of the message
