Arg. TCP slow start killing me.

[Jason Wolfe]

Erich,

Slow start is actually just the initial ramp up limited by RFC 3390 being
enabled by default (usually 3/4 packets), but this is only the case for the
first few seconds of the stream.  You can effectively speed that up with
something like this though:

net.inet.tcp.rfc3390=0
net.inet.tcp.slowstart_flightsize=10
net.inet.tcp.sendspace=262144
net.inet.tcp.recvspace=262144

The first 2 allow 10 packets to be sent before an ACK, and the 2nd 2 just
bump as the starting window size.  With your memory and the massive max you
set no reason to force them to slowly step up from such a low initial size.
 Looks like the numbers you used for initial are actually the default
increment/step size of the window growth.

Also since you mentioned latency playing a factor here, try this sysctl.
 If overruns are an issue you'll likely see a bit of an increase in
retransmits, but could potentially show a sizable positive impact in the
saw tooth.

net.inet.tcp.inflight.enable=0

Is it possible to upgrade to 8.2-STABLE?  Cubic has shown some really great
improvement in my latent paths, a steady 10% overall increase in same cases.

Jason Wolfe

On Sun, Nov 13, 2011 at 2:16 PM, Erich Weiler <weiler at soe.ucsc.edu> wrote:

> So, I have a FreeBSD 8.1 box that I'm using as a firewall (pfSense 2.0
> really, which uses 8.1 as a base), and I'm filtering packets inbound and
> I'm seeing a typical sawtooth pattern where I get high bandwidth, then a
> packet drops somewhere, and the TCP connections back off a *lot*, then
> slowly get faster, then backoff, etc.  These are all higher latency WAN
> connections.
>
> I get an average of 1.5 - 2.0 Gb/s incoming, but I see it spike to like
> 3Gb/s every once in a while, then drop again.  I'm trying to maintain that
> 3Gb/s for as long as possible between it dropping.
>
> Given that 8.1 does not have the more advanced TCP congestion algorithms
> like cubic and H-TPC that might help that to some degree, I'm trying to
> "fake it".  ;)
>
> My box has 24GB RAM on it.  Is there some tunable I can set that would
> effectively buffer incoming packets, even though the buffers would
> eventually fill up, just to "delay" the TCP dropped packet signal telling
> the hosts on the internet to back off?  Like, could I effectively buffer
> 10GB of packets in the queue before it sent the backoff signal?  Would
> setting kern.ipc.nmbclusters or something similar help?
>
> Right now I have:
>
> loader.conf.local:
>
> vm.kmem_size_max=12G
> vm.kmem_size=10G
>
> sysctl.conf:
>
> kern.ipc.maxsockbuf=16777216
> kern.ipc.nmbclusters=262144
> net.inet.tcp.recvbuf_max=16777216
> net.inet.tcp.recvspace=8192
> net.inet.tcp.sendbuf_max=16777216
> net.inet.tcp.sendspace=16384
>
> I guess the goal is to keep the bandwidth high without dropoffs for as
> long as possible, with out as many TCP resets on the streams.
>
> Any help much appreciated!  I'm probably missing a key point, but that's
> why I'm posting to the list.  ;)
>
> cheers,
> erich
>
> _______________________________________________
> freebsd-net at freebsd.org mailing list
> http://lists.freebsd.org/mailman/listinfo/freebsd-net
> To unsubscribe, send any mail to "freebsd-net-unsubscribe at freebsd.org"
>