1. Jiyong Park Seoul National University, Korea 2005. 2. 22.
Preventing network performance interference of bidirectional TCP traffic with packet scheduling in a home network gateway using an asymmetric link
Jiyong Park
Seoul National University, Korea

2. Contents Overview of TCP Network model
Performance problems of TCP on asymmetric link
Causes
Normalized asymmetry ratio
Bottleneck buffer size
Competing reverse traffic
Proposed Solutions
CBQ
CBQ with simple implementation
ACK filtering (AF) / ACK reconstruction (AR)
CBQ+AF/AR
ACQ+AF/AR
Our problem

3. Related Papers
T.V. Lakshmnan et al., Window-based error recovery and flow control with a slow acknowledgement channel: a study of TCP/IP performance, INFOCOM, (1)
L. Kalampoukas et al., Improving TCP Throughput over Two-Way Asymmetric Links: Analysis and Solutions, SIGMETRICS, (2)
H. Balakrishnan et al., The Effects of Asymmetry on TCP Performance, MobiCom, (3)

4. TCP Overview Three key mechanisms of TCP Reliability
Acknowledge
Flow control (speed difference between sender and receiver application)
Sliding window
Window advertisement
Congestion control (packet loss, delay, ...)
Slow start
Congestion avoidance

5. TCP Overview – 2. Flow Control
Window
Number of packets that can be sent without acknowledged
Sliding window
Upon receiving ACK, window slides
Window advertisement
Receiver advertises its window size on ACK (sender cannot send more than that)

6. TCP Overview – 2. Flow Control
Sender Buffer
Sent ACKed
Sent Not ACKed
Will send
TCP output
window (size 5)
user output 1 2 3 4 5 6 7 8 9 10 11 11 12 12 13 14 13 14
sequence number
ACK received: 6 9

7. TCP Overview – 3. Congestion Control
Two kinds of windows
Congestion window (cwnd)
Advertised window (awnd)
actual window size = min(cwnd, awnd)
Upon receiving ACK, increases cwnd (try to send more)
Slow start (aggressive) (cwnd++ upon receiving 1 ACK)
Congestion avoidance (cautious) (cwnd ++ upon receiving cwnd ACKs)
Congestion is caused by
Data packet loss (detected by duplicated ACK)
Long delay (detected by timer expiration)
Upon detecting congestion, decreases cwnd = 1 (restart)

8. TCP Overview – 3. Congestion Control
Enters congestion avoidance phase when W > Wt
Wt := Wt / 2 when congestion is detected
congestion window size
congestion detected
time

9. TCP Overview – Packet Processing
TCP protocol handler performs two major jobs
Upon receiving ACK, send data packet
Upon receiving data, deliver it to user and send ACK
These two jobs are handled by one handler (no interleaving)
User process
Data packet
IP(input) queue
Output queue
data
ACK
ACK packet
ACK
data
awaken by interrupt handler
Protocol handler

10. Network Model (1/2)
BS : 66
US : 800Kbps
Local Host
TS : 10ms
Remote Host
TF: 10ms
UF : 2.1Mbps
BF : 66
WL: 66
WR : 66
W (maximum window size): maximum number of packets that can be in the network
U: transmission speed in bps
B: queue size in packets
T: propagation delay
Data packet size: 1040 byte
ACK packet size: 40 byte

11. Network Model (2/2) Application programs are assumed to be greedy.
There are always packets to send.
There are always rooms to receive packets.

12. Asymmetry Ratio (1/2) Asymmetry Ratio (k)
Service speed for data packet on forward link / Service speed for ACK packet on reverse link
Example
Fast link can send 1000 data packets per second
Slow link can send 25 ACK packets per second
Then k = 1000/25 = 40

13. Asymmetry Ratio (2/2)
At the receiving side, one ACK is generated for one data packet received.
If k > 1, ACKs are queued.
If k <= 1, no problem.
1000 data pkt/s
25 ACK pkts/s
Local Host
Remote Host
ACK
Data
1000 data pkt/s

14. Lets first consider the case when only download traffic is present.

15. sequence number of data
Effect of Small k
ACK arrives periodically and smoothly (sequence number increases by 1)
Smooth transmission
sequence number
sequence number of data
sequence number of ACK
time

16. Effect of Large k Output buffer size of slow link is the key.
Size is too small: (k-1) out of k ACKs are dropped
Size is too large: entire ACKs are queued
4m pkt/s
m pkt/s 9 5 1 6 2 7 3
dropped 8 4 9 8 7 6 5 4 3 2 1

17. Large k with Small Buffer (1/2)
Since ACK is cumulative, sequence number increases abruptly by k
 Bursty transmission (output buffer is needed)
sequence number
sequence number of data
k burst
sequence number of ACK
k burst
time

18. Large k with Small Buffer (2/2)
Output buffer on fast link should larger than k
Bursty packets are regulated by buffer
queue size
burst
regulated
time

19. Large k with Large Buffer
Large buffer is problematic.
Throughput is decreased.
ACKs are very delayed.
Sender has to wait to receive the ACKs.
Idle time occurs.
Example:
Maximum window size: 100 packets.
k = 10
Time to send full window of data packets: t seconds
Time to send full window of ACK packets: 10t seconds
RTT = 11t seconds.
Achieved throughput = 100 / 11t = 9/t
Ideal throughput = 100/t

20. Summary of Single Download Connection
k <= 1
No problem (our case)
k > 1
Size of output buffer on slow link is small
Bursty transmission.
Solved by using output buffer on fast link (size > k)
Size of output buffer on slow link is very big
Big problem.
Very long delay of ACKs. Reduced throughput.

21. Now, proceed to the case of two-way traffic.

22. Two-way Connection On uplink,
ACK packets for download traffic
Data packets for upload traffic
Size of output buffer on the link is small
Upload speed is reduced.
Size of output buffer on the link is large
Download speed is reduced.
Local Host
ACK
Data
Case of our ADSL Modem

23. Two-way Connection – Case 1
Size of output buffer on slow link is small
Upload speed is reduced.
ACK and data packets are also dropped.
Dropping ACK is not a big problem (solvable by buffer)
 Download speed is not affected.
Dropping data packet is a big problem
Timeout  Retransmission
 Upload speed is reduced.

24. Two-way Connection – Case 2
Size of output buffer on slow link is large
Download speed is reduced.
ACKs are queued behind the big data packets
ACK compressing
ACK packets for download
data packets for upload 9 8 7 6 5 4 3 2 1 9 8 7 6 5 4 3 2 1
Very big delay

25. Two-way Connection - Our Case
k = 0.1 (ACK is seldom dropped)
Window size = 66
Upload buffer size = 100
Upload buffer is small
 Upload speed is reduced.
Buffer size : 100
34 data packets for upload traffic
66 ACKs for download traffic
23 data packets are dropped

26. Proposed Solutions Paper #1 Paper #2 Paper #3
Upload buffer size is small
Upload throughput is reduced
Solution: round-robin queue scheduling
Paper #2
Upload buffer size is big
Download throughput is reduced
Solution: queue scheduling
Paper #3
Solution: ACC, AF, ACK-first, Sender adaptation
Similar to our approach (More flexible than us)
Similar to our approach
(Better performance than us)
We can adopt some techniques from here

27. Solution in the Paper #1 Round-robin queue scheduling f is big
ACK for download 9 8 7 6 5 4 3 2 1 f
RR scheduler
data for upload 9 8 7 6 5 4 3 2 1
1-f
f is big
Upload: slow (starvation)
Download: good
f is small
Upload: good (1-f of the full speed)
Download: slowdown (bursty transmission and small size of fast buffer)

28. Solution in the Paper #2
Emulate fair queuing scheduler by counter-based implementation
Upload speed is guaranteed by (1-f) of the full speed.
Download speed is reduced
Slow buffer size is large
Uses simple FIFO
ACKs are delayed very long

29. Solution in the Paper #3 Ack Filtering (AF) queuing
Ack-first scheduling AF
Remove all redundant cumulative ACKs in the queue
If there is an ACK packet in the queue, always transmit it first. 10 9 8 7 6 5 4 3 2 1

30. Solution in the Paper #3 Download: full speed (no delay of ACK)
Upload: starvation
ACKs are arrived faster than they are processed.
So, there are always at least one ACK in the queue.
ACK has strictly higher priority than data.
Data has no chance to be transmitted.  starvation

31. My Previous Solution Solution: Give higher priority to ACK over data
Result
Other papers: download performance was bad.
Our result: both traffic was good.
Reason: k is very small in our environment.
Bandwidth consumed by ACK packets is very low
About 11% (94Kbps on 800Kbps link)
Enough bandwidth to send data packet for upload traffic

32. Our New Solution AF + Queue scheduling Queue scheduling AF
Guarantees upload throughput AF
Minimizes ACK latency for download traffic

33. Better Solution Proposed (1/5)
Fatma Louati, Chadi Barakat, Walid Dabbous, Handling two-way TCP traffic in asymmetric networks, Speed Networks and Multimedia Communications (HSNMC), 2004
AF + AR + ACQ

34. Better Solution Proposed (2/5)
AF + AR + ACQ AF
Minimizes ACK latency for download traffic
AR (ACK Regeneration)
Regulates burst transmission caused by AF
ACQ (Adaptive Class-based Queuing)
A kind of CBQ in which the bandwidth ratio is changed automatically and periodically
RFC 3449 recommendations (by the authors of the paper #3)

35. Better Solution Proposed (3/5)
Ratio update algorithm
rn: bandwidth allocated to data packet of upload traffic
current download speed
current upload speed

36. Better Solution Proposed (4/5)
Performance metric
Link efficiency of upload traffic + link efficiency of download traffic
from 0 to 2
ACQ+AF/AR : 1.72

37. Better Solution Proposed (5/5)
strong asymmetry
weak asymmetry