1. TCP Westwood
End-to-End Congestion Control for Wired/Wireless Networks
By
Guillem C. Mayol
Adrian Viala
Dídac Montero
Mateu Seguí

2. TCP Westwood: Main features
¤ Improves the performances of TCP Reno over leaky and
dynamically loaded canal as wireless networks (small
improve for wired network)
¤ TCPW in not very sensitive to random errors.
¤ TCPW fully complies with end-to-end TCP design principle.
¤ Not requires inspection of TCP packets at intermediate nodes.
¤ BW estimated by continuously monitoring ACKs.
¤ Friendliness with TCP variants, such as Reno.
¤ TCPW is a reactive process.
¤ TCPW convergence to fair share.

3. TCP Westwood: How works?
¤ TCPW works as TCP Reno but with smallest changes.
¤ After 3 ACK losses
¤ ssthresh = (BWE * RTTmin)/seg_size;
¤ If (cwin > ssthresh) cwin = ssthresh;
¤ Timeout expiration
¤ ssthresh = (BWE * RTTmin)/seg_size; (min 2) 3 ACK losses
¤ cwin = 1;
Timeout

4. TCP Westwood: How works?
Image source: http://c3lab.poliba.it/index.php/Westwood

5. delivered to the destination. equal to τ/2 is necessary. But, since the ACK
We discuss the use of the information in (2) in section 2.3. chronous, the sampling frequency constraint c
For now, let assume that an ACK is received at the source at anteed. Thus, to guarantee the Nyquist constr
time tk , notifying that dk bytes have been received at the TCP lish that if a time τ/m (m 2) has elapsed
receiver. We can measure the following sample bandwidth received ACK without receiving any new AC
used by that connection as bk = dk / k , where k = tk −tk−1 ter assumes the reception of a virtual null sa
and tk−1 is the time the previous ACK was received. The situation is shown in figure 1, where tk
Since congestion occurs whenever the low-frequency input an ACK is received, tk+j are the arrival time
TCP Westwood: End-to-End
traffic rate exceeds the link capacity [15] we employ a low- samples, with tk+j +1 − tk+j = τ/m for j
pass filter to average sampled measurements and to obtain the and bk+j = 0 for j = 0, n − 1 are th
low-frequency components of the available bandwidth. No- ples. Then, bk+n = dk+n / k+n is the bandw
bandwidth measurement I
tice that this averaging is also useful in filtering out the noise tk+n .
due to delayed acknowledgments. It is desirable that after a long time witho
In our early design and experimentation, we used a filter because no new data were sent), the filter ac
¤ BW estimated (BWE) by monitoring the TCP ACKs
similar to the one used for RTT estimation in TCP. We de- vative fashion, progressively decreasing the b
termined that such an exponential filter with constant coeffi- mation as time elapses without new samples.
cients is not capable of efficiently filtering out high-frequency the operation of the TCPW filter when there is
Transmitted Bytes
components of the bandwidth measurements. We propose the sence of ACKs after a time t = tk . As can be
dk dk
bk = = ;
BW
Δ k tk − tk−1
Interarrival ACKs ACK received at source

6. delivered to the destination. equal to τ/2 is necessary. But, since the ACK
We discuss the use of the information in (2) in section 2.3. chronous, the sampling frequency constraint c
For now, let assume that an ACK is received at the source at anteed. Thus, to guarantee the Nyquist constr
time tk , notifying that dk bytes have been received at the TCP lish that if a time τ/m (m 2) has elapsed
receiver. We can measure the following sample bandwidth received ACK without receiving any new AC
used by that connection as bk = dk / k , where k = tk −tk−1 ter assumes the reception of a virtual null sa
and tk−1 is the time the previous ACK was received. The situation is shown in figure 1, where tk
Since congestion occurs whenever the low-frequency input an ACK is received, tk+j are the arrival time
TCP Westwood: End-to-End
traffic rate exceeds the link capacity [15] we employ a low- samples, with tk+j +1 − tk+j = τ/m for j
pass filter to average sampled measurements and to obtain the and bk+j = 0 for j = 0, n − 1 are th
low-frequency components of the available bandwidth. No- ples. Then, bk+n = dk+n / k+n is the bandw
bandwidth measurement II
tice that this averaging is also useful in filtering out the noise tk+n .
due to delayed acknowledgments. It is desirable that after a long time witho
In our early design and experimentation, we used a filter because no new data were sent), the filter ac
similar to the one used for RTT estimation in TCP. We de- vative fashion, progressively decreasing the b
termined that such an exponential filter with constant coeffi- mation as time elapses without new samples.
¤ TCPW uses a low pass filter to average sampled
cients is not capable of efficiently filtering out high-frequency
components of the bandwidth measurements. We propose the
the operation of the TCPW filter when there is
sence of ACKs after a time t = tk . As can be
measurements and to obtain low-freq. components of the
available bandwidth. BW (Actual) ACK received at source
BWE
ˆ = α b + (1− α )( bk + bk−1 ) α k = 2τ − Δ k
bk ˆ
k k−1 k
2 2τ + Δ k
Last BWE
τ =1/cut-off freq. Interarrival ACKs
¤ When interarrival ACKs increases (losses?) the most
important values are the two most recent BW calculated,
otherwise the Last BWE has more weight.

7. TCP Westwood: End-to-End
bandwidth measurement III

8. TCP Westwood: End-to-End
bandwidth measurement IV

9. TCP Westwood: performance with
lossy link & fair share
¤ Avg. throughput versus
number of Reno connections
over good and lossy link.
" Convergence toward fair
bandwidth sharing when
connection A started firstly

10. TCP Westwood: performance

11. TCP Westwood: performance