rwnd autotuning & cwnd growth
Cheng.Cui at netapp.com
Wed Apr 4 19:34:15 UTC 2018
On 4/4/18, 1:19 PM, "owner-freebsd-transport at freebsd.org on behalf of Scheffenegger, Richard" <owner-freebsd-transport at freebsd.org on behalf of Richard.Scheffenegger at netapp.com> wrote:
Wasn't there a burst limit included in freebsd tcp at some point in time?
Today I observed the following scenario: client window host, with receive window autotuning enabled; server freebsd 10.4; latency around 20ms (bufferbloated; unloaded latency around 1-2ms), the link stretches about 5 links with 10G,25G,40G each (server at 10G), client is 100Mbit.
The client signals for the longest time a receive window of around 350kB, which the server hogs for transmitting data (effective throughput < 100Mbit), after growth of cwnd hits rwnd. Then, after a couple seconds, the client signals an increase of the receive window by a factor of ~2 (700kB), about where the cwnd has grown in the background if the transmission rate weren't clamped by rwnd.
Thus the server sends a line rate (10G) burst of about 300kB (cwnd has not increased to completely include the new rwnd), which is queued up (and the tail dropped...), increasing the effective latency to about 40ms...
Just looked at rfc5681, 2861 - but apparently receive window autoscaling is too new a technology for those;
Linux does have a burst clamping heuristic implemented, to prevent humongous burst of data to be sent, when for whatever reason cwnd would allow it (rwnd clamping and increasing, application limitated operation, ...) But that apparently never made it into any spec?
IMHO, limiting the maximum burst size per RTT, or doing slow start when cwnd suddenly allows for an "excessive" amount of data to be sent, would be the sensible thing to do - but whats your view on this?
freebsd-transport at freebsd.org mailing list
To unsubscribe, send any mail to "freebsd-transport-unsubscribe at freebsd.org"
More information about the freebsd-transport