DTracing TCP congestion control

In a previous post, I showed how we can use DTrace to probe TCP receive and send window events. TCP receive and send windows are in effect both about flow-controlling how much data can be received - the receive window reflects how much data the local TCP is prepared to receive, while the send window simply reflects the size of the receive window of the peer TCP. Both then represent flow control as imposed by the receiver.

However, consider that without the sender imposing flow control, and a slow link to a peer, TCP will simply fill up it's window with sent segments. Dealing with multiple TCP implementations filling their peer TCP's receive windows in this manner, busy intermediate routers may drop some of these segments, leading to timeout and retransmission, which may again lead to drops. This is termed congestion, and TCP has multiple congestion control strategies. We can see that in this example, we need to have some way of adjusting how much data we send depending on how quickly we receive acknowledgement - if we get ACKs quickly, we can safely send more segments, but if acknowledgements come slowly, we should proceed with more caution. More generally, we need to implement flow control on the send side also.

Slow Start and Congestion Avoidance

From RFC2581, let's examine the relevant variables:

"The congestion window (cwnd) is a sender-side limit on the amount of data the sender can transmit into the network before receiving an acknowledgment (ACK). Another state variable, the slow start threshold (ssthresh), is used to determine whether the slow start or congestion avoidance algorithm is used to control data transmission"

Slow start is used to probe the network's ability to handle transmission bursts both when a connection is first created and when retransmission timers fire. The latter case is important, as the fact that we have effectively lost TCP data acts as a motivator for re-probing how much data the network can handle from the sending TCP. The congestion window (cwnd) is initialized to a relatively small value, generally a low multiple of the sending maximum segment size. When slow start kicks in, we will only send that number of bytes before waiting for acknowledgement. When acknowledgements are received, the congestion window is increased in size until cwnd reaches the slow start threshold ssthresh value. For most congestion control algorithms the window increases exponentially under slow start, assuming we receive acknowledgements. We send 1 segment, receive an ACK, increase the cwnd by 1 MSS to 2*MSS, send 2 segments, receive 2 ACKs, increase the cwnd by 2*MSS to 4*MSS, send 4 segments etc. When the congestion window exceeds the slow start threshold, congestion avoidance is used instead of slow start.

During congestion avoidance, the congestion window is generally updated by one MSS for each round-trip-time as opposed to each ACK, and so cwnd growth is linear instead of exponential (we may receive multiple ACKs within a single RTT). This continues until congestion is detected. If a retransmit timer fires, congestion is assumed and the ssthresh value is reset. It is reset to a fraction of the number of bytes outstanding (unacknowledged) in the network. At the same time the congestion window is reset to a single max segment size. Thus, we initiate slow start until we start receiving acknowledgements again, at which point we can eventually flip over to congestion avoidance when cwnd > ssthresh.

Congestion control algorithms differ most in how they handle the other indication of congestion - duplicate ACKs. A duplicate ACK is a strong indication that data has been lost, since they often come from a receiver explicitly asking for a retransmission. In some cases, a duplicate ACK may be generated at the receiver as a result of packets arriving out-of-order, so it is sensible to wait for multiple duplicate ACKs before assuming packet loss rather than out-of-order delivery. This is termed fast retransmit (i.e. retransmit without waiting for the retransmission timer to expire). Note that on Oracle Solaris 11, the congestion control method used can be customized. See here for more details. In general, 3 or more duplicate ACKs indicate packet loss and should trigger fast retransmit . It's best not to revert to slow start in this case, as the fact that the receiver knew it was missing data suggests it has received data with a higher sequence number, so we know traffic is still flowing. Falling back to slow start would be excessive therefore, so fast recovery is used instead.

Observing slow start and congestion avoidance

The following script counts TCP segments sent when under slow start (cwnd <= ssthresh) versus congestion avoidance (cwnd > ssthresh).

#!/usr/sbin/dtrace -s

#pragma D option quiet

tcp:::connect-request
/ start[args[1]->cs_cid] == 0/
{
        start[args[1]->cs_cid] = 1;
}
tcp:::send
/ start[args[1]->cs_cid] == 1 &&
  args[3]->tcps_cwnd <= args[3]->tcps_cwnd_ssthresh /
{
        @c["Slow start", args[2]->ip_daddr, args[4]->tcp_dport] = count();
}

tcp:::send
/ start[args[1]->cs_cid] == 1 &&
  args[3]->tcps_cwnd > args[3]->tcps_cwnd_ssthresh /
{
        @c["Congestion avoidance", args[2]->ip_daddr, args[4]->tcp_dport] =
            count();
}

As we can see the script only works on connections initiated since it is started (using the start[] associative array with the connection ID as index to set whether it's a new connection (start[cid] = 1). From there we simply differentiate send events where cwnd <= ssthresh (slow start) versus those where cwnd > ssthresh (congestion avoidance). Here's the output taken when I accessed a YouTube video (where rport is 80) and from an FTP session where I put a large file onto a remote system.

# dtrace -s tcp_slow_start.d
^C
ALGORITHM                 RADDR                     RPORT      #SEG
Slow start                10.153.125.222               20         6
Slow start                138.3.237.7                  80        14
Slow start                10.153.125.222               21        18
Congestion avoidance      10.153.125.222               20      1164

We see that in the case of the YouTube video, slow start was exclusively used. Most of the segments we sent in that case were likely ACKs. Compare this case - where 14 segments were sent using slow start - to the FTP case, where only 6 segments were sent before we switched to congestion avoidance for 1164 segments. In the case of the FTP session, the FTP data on port 20 was predominantly sent with congestion avoidance in operation, while the FTP session relied exclusively on slow start.

For the default congestion control algorithm - "newreno" - on Solaris 11, slow start will increase the cwnd by 1 MSS for every acknowledgement received, and by 1 MSS for each RTT in congestion avoidance mode. Different pluggable congestion control algorithms operate slightly differently. For example "highspeed" will update the slow start cwnd by the number of bytes ACKed rather than the MSS.

And to finish, here's a neat oneliner to visually display the distribution of congestion window values for all TCP connections to a given remote port using a quantization. In this example, only port 80 is in use and we see the majority of cwnd values for that port are in the 4096-8191 range.

# dtrace -n 'tcp:::send { @q[args[4]->tcp_dport] = quantize(args[3]->tcps_cwnd); }'
dtrace: description 'tcp:::send ' matched 10 probes
^C

    80
           value  ------------- Distribution ------------- count    
              -1 |                                         0        
               0 |@@@@@@                                   5        
               1 |                                         0        
               2 |                                         0        
               4 |                                         0        
               8 |                                         0        
              16 |                                         0        
              32 |                                         0        
              64 |                                         0        
             128 |                                         0        
             256 |                                         0        
             512 |                                         0        
            1024 |                                         0        
            2048 |@@@@@@@@@                                8        
            4096 |@@@@@@@@@@@@@@@@@@@@@@@@@@               23       
            8192 |                                         0        

Comments:

Post a Comment:
Comments are closed for this entry.
About

user12820842

Search

Archives
« April 2014
SunMonTueWedThuFriSat
  
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
   
       
Today