On 09/03/15 at 09:13P, Lawrence Stewart wrote: > On 09/03/15 10:54, hiren panchasara wrote: > > I am failing to understand the reason behind this behavior. > > > > What should the congestion window (snd_cwnd) be set to when we hit loss? > > It seems that we set it to 1 segment right now. > > https://svnweb.freebsd.org/base/head/sys/netinet/tcp_input.c?revision=286227&view=markup#l2531 > > > > I also see that in the simulations I did. Sender side pcap can be found > > at: https://people.freebsd.org/~hiren/pcaps/single_packet_loss.pcap > > > > Trying to send 50kb of data from freebsd 10.2 server to freebsd client. > > Initial cwnd is 10 so we blast out 10 packets but 1 packet gets dropped: > > seq 2897:4345. We get 3 dupacks and we retransmit it. But as soon as we > > detect this loss, we reduce cwnd to 1 segment. In fact, we could've used > > data in SACK to see how much we could send on the n/w, imo. > > > > 3rd dup ack (which triggered the retransmit) looks like this: > > IP 192.168.11.10.41674 > 192.168.10.10.http: Flags [.], ack 2897, win > > 12579, options [nop,nop,TS val 4236220288 ecr 3905376863,nop,nop,sack 1 > > {4345:10137}], length 0 > > > > And the retransmit: > > IP 192.168.10.10.http > 192.168.11.10.41674: Flags [.], seq 2897:4345, > > ack 172, win 12579, options [nop,nop,TS val 3905376894 ecr 4236220288], > > length 1448 > > > > At this point in time, sender knows that it has sent 23169 bytes (last > > packet server sent was seq 21721:23169) and received ack for 10137 > > bytes minus a missing packet = 8689 bytes. i.e. 6 packets. So, there is > > at least that much room on n/w at that point in time. We can go > > conservative and halve that. i.e. 3 packets. That is still better than > > going down to 1 packet. > > > > Is there something basic I am missing here? > > Any insights would be helpful. > > You want to read up about window inflation during fast recovery in RFC > 5681 followed by 3782, and then consult Stevens vol 2 to understand how > variables are used for different purposes depending on connection state > and which code path was taken (something I greatly dislike and would > love to change one day).
Thanks Lawrence for the pointers. From what I could see in the code (as I pointed above in original email) as soon as we get 3rd dup ack in tcp_do_segment() insdie 'else if (tp->t_dupacks == tcprexmtthresh)' we check if we could do SACK with 'if (tp->t_flags & TF_SACK_PERMIT)' and we set 'tp->snd_cwnd = tp->t_maxseg;' before calling tcp_output(). And I've seen this behavior on many situations. So wanted to confirm if others are also seeing this. FWIW, I am using cubic cc. But as far as I can see, snd_cwnd setting for FreeBSD is outside of any cc algos' scope. i.e. We override whatever is being set. I'll do the necessary readings. Cheers, Hiren
pgpNpYJgalKhn.pgp
Description: PGP signature
