TCP Africa Summary

TCP Africa An Adaptive and Fair Rapid Increase Rule for Scalable TCP

Sai Deep 2012CS10223

1

Overview TCP - Review Design Considerations for a high speed protocol Loss Based vs Delay Based TCP TCP Africa Experimental Study Summary

2

Transmission Control Protocol Connection oriented

Three-Way Handshake

Reliable, in-order delivery ACKs, retransmissions, sequence numbers, checksums

Congestion Control Reduce transmission rate when congestion occurs

Flow Control Sender does not overwhelm receiver

3

TCP Reno Most deployed protocol in the Internet Additive Increase Multiplicative Decrease Congestion Avoidance Phase (cwnd ssthresh)

On each successful ACK cwnd cwnd + 1/cwnd

On triple duplicate ACKs cwnd cwnd/2

4

5

High Speed Scenario Desired congestion window is quite high

Roughly equal to the BDP of the connection ~83,000 packets for a 10 Gbps link with a 100 ms RTT

Requires a lot of time for window to be regained after a loss Low utilization even when the network is uncongested

6

Design Considerations for a high speed protocol Throughput

Efficient utilization of high available bandwidth Peer fairness

Fairness between two flows of differing RTTs

TCP-fairness Should be fair with the older TCP-Reno standard

Congestion collapse Major concern since they are more aggressive than TCP-Reno

7

STCP Most aggressive of the current well known TCP proposals Multiplicative Increase Multiplicative Decrease

On each successful ACK cwnd cwnd + 0.01 * cwnd

On a packet drop cwnd 0.875 * cwnd

8

HSTCP Scales its drop parameter from 50% at low window sizes to

90% at higher window On each successful ACK

w w + a(w) / w On triple duplicate ACKs

w w - b(w) * w a(w) and b(w) are functions of the current window size

9

10

Loss based high speed protocols Excellent Scalability

Rapid return to the maximum window At the cost of frequent self induced congestion events STCP - every 13.4 RTTs (regardless of the link speed)

Both have undesirable fairness properties Poor RTT bias Suppress TCP-Reno

11

Delay based TCP - FAST TCP Round trip delay as signal for congestion Unlikely to cause significant queuing delay Can quickly converge to equilibrium Can run in steady state without causing packet drops Disadvantage

Cant compete with TCP-Reno Primary reason preventing its widespread adoption

12

TCP Africa Hybrid protocol Uses a delay metric to determine congestion Operates in two modes

FAST mode - in the absence of congestion aggressive congestion avoidance rule of HSTCP

SLOW mode - in the presence of congestion conservative Reno congestion avoidance rule

13

Delay metric Based on TCP-Vegas

aRTT Exponentially smoothed high accuracy RTT estimateminRTT minimum delay observed on the path a constant, usually a real number larger than one

14

Algorithmif ( aRTT minRTT < * aRTT/W)

W = W + fast_increase(W)/W else

W = W + 1/W Flows with small RTT do not gain a competitive advantage

Improved RTT bias performance

15

Experimental Study Comparison of the performances of TCP-Africa & HSTCP parameter set to 1.65

16

Safety Investigate if a protocol hampers the performance of other

flows

17

622 Mbps 80ms delay

1 Gbps 1ms delay

100 Mbps 1ms delay

Results HSTCP had a significant effect on the Reno throughput Ratio of HSTCP to Reno traffic roughly 25:1 2700 non slow-start packets lost at the bottleneck link TCP-Africa had a minimal effect on the Reno flow Ratio of TCP-Africa to TCP-Reno roughly 6:1 Only 47 non slow-start related packets were lost

18

Fairness with TCP-Reno

21

622 Mbps 80ms delay

1 Gbps 1ms delay

1 Gbps 1ms delay

Results Ratio of HSTCP to TCP-Reno traffic 17:1 2000 non slow-start packets lost at the bottleneck link Ratio of TCP-Africa to TCP-Reno roughly 4:1 Only 32 non slow-start related packets were lost

22

RTT Bias Flow from A-B has 30ms RTT Flow from C-D is set such that its RTT is a multiple of 30ms HSTCP

Serious fairness problem with flows of different RTT Short RTT flow quickly dominates the connection

Starves the other flow TCP-Africa

Flows share the bandwidth proportional to their RTTs25

26

Adapting to network conditions Study how quickly TCP-Africa can adapt to changing

network conditions The bottleneck link has a capacity of 622 Mbps Flow experiencing a minimum RTT of 84 ms After 160s, a UDP flow at 300 Mbps is started At 320s, the UDP flow is stopped

27

Results Quickly reduces its bandwidth in response to the UDP flow As the flow approaches the maximum available bandwidth

Enters slow mode After the UDP flow terminates

Quickly re-enters high speed mode Quickly utilizes the newly freed bandwidth

28

29

Summary Maintaining a careful balance between aggressiveness, fairness & safety

A major challenge in developing high speed protocols TCP-Africa exploits congestion indicators towards fair rapid increase

Not giving in to Reno, yet not crushing it In the experiments, it achieved

Excellent utilization of bandwidth Low induced packet loss rate Excellent fairness properties, RTT bias performance

Overall, a good transfer protocol model Has the potential to be used in the future

30

Thank You