AIRA: Additive Increase Rate Accelerator

Ioannis Psaras, Vassilis Tsaoussidis

Democritus University of ThraceXanthi, Greece

AIRA: an overview

• Congestion Control: A classic problem with no satisfactory solution – simplicity dominated.

• We propose– A new model to control congestion

• Graft an adjustable acceleration capacity to network flows

– Rules to determine increase/decrease policy• Avoid flow starvation

• Maintain equilibrium

• Exploit resources fast or slow, depending on contention

Motivating Arguments• Congestion events in modern networks are rather

transient events – not persistent as in the past

• Congestion remains the problem and the goal – we identify a network state based on feedback from congestion

• Congestion is, presently, not directly associated with contention – two flows can cause congestion.

• Congestion window only determines an initial state – does not tell the whole story for the day after.

• Rate increase determines a policy based on some initial condition – does not cope with network dynamics

To get to the point…

• The fixed Additive Increase factor of AIMD is similar to a car regulating its velocity scale but with fixed acceleration.

• Two flows or two hundred flows can cause congestion. Their response is similar after congestion – is that not a problem?

– Although AIMD-systems are adaptive to network dynamics, this adaptivity is limited: the cwnd size can be regulated, but the cwnd increase rate (i.e., AI factor) cannot.

Contribution

• We introduce a new congestion control paradigm for the responsive behavior of flows, when bandwidth availability changes due to varying network contention.

• The Additive Increase Rate Accelerator (AIRA) adjusts (progressively) the Additive Increase factor of AIMD according to the current level of network contention.

Contribution Framework

• Congestion control should incorporate contention into the responsive strategy– Flows should apply rates that

correspond to the level of contention

• Contention has to be estimated, at least coarsely– Protocol effort/gain evolution can be

measured and determine contention increase/decrease

Adaptive rates

• Therefore, we rely on two major concepts to move beyond the confined perspective of predetermined and inflexible congestion control:

• (i) effort-based contention estimation • (ii) contention-oriented adaptive Additive

Increase transmission.

Design Perspective

• Estimate contention

• Calculate increase/decrease rules for increase rates– Efficiency=>speedy increases when contention

is low, slow responses when contention is high

• Calculate (each flow) operational rounds and incorporate duration into rate increase policy – this preserves fairness for short-lived flows

• Guarantee that flows will not starve

JustificationOne possible justification for not highlighting that research direction is that:The Additive Increase factor of AIMD does not contribute to the long-term Goodput performance of TCP.

Justification

The area underneath the solid cwnd lineplot (Area 1 and 2) represents the Goodput performance of the protocols.

Therefore, both protocols achieve the same Goodput performance, since A1=A2 and A3=A4

Justification: Goodput, EffortHowever,• Timeouts are not accounted in the graph –

more congestion events deteriorate system performance further

• Additive Increase affects significantly the Retransmission Effort of flows, which impacts overall system behavior, as well.– For example, TCP (a = 1), experiences 4

congestion events, while TCP (a = 0.5) experiences only 2. Assuming that each congestion event is associated with a fixed number of lost packets, regular TCP (i.e., a = 1) will retransmit twice as many packets as TCP with a = 0.5, without any gain in Goodput performance.

How can we estimate contention?

• Keep track of effort (throughput) and gains (goodput)

• When the gap increases within a round, without having the flow behavior changing, contention has been increased

• Otherwise, contention has been decreased

System Model: Definitions1. A Round is defined as the interval between two cwnd

multiplicative decreases.2. Round Loss Rate (p_i) is the ratio of the lost packets

over the total number of sent packets, within Round i. The Round Loss Rate is calculated at the end of each Round.

3. The Throughput Slope within a Round is defined as a/cwnd. Obviously, the Throughput Slope is identical to the cwnd Slope.

4. Assuming a Round Loss Rate p_i and a Round duration t_i, the Desired Throughput Slope within a Round is defined as the hypothetical cwnd Slope, which would result in zero packet losses, within t_i, but without causing bandwidth underutilization. The Desired Throughput Slope is, therefore, determined by:

∙(1 )ia p

cwnd−

System Model: Definitions

System Model: Definitions

• When p_i > p_{i-1} then the Throughput Slope exceeds the Desired Throughput Slope, for Round i.– The greater the distance between p_i and p_{i-

1}, the wider the gap between the Throughput and Desired Throughput Slopes.

• When p_i < p_{i-1} then the Throughput Slope is underneath the Desired Throughput Slope, for Round i.– The greater the distance between p_i and p_{i-

1}, the wider the gap between the Desired Throughput and Throughput Slopes.

System Model: Solution Framework

1. Converge to Fairness and ALPHA Fairness:

2. Guarantee Stability

3. Exploit Available resources efficiently:– Faster (i.e., aggressively), when contention is low

– Slower (i.e., conservatively), when contention is high

Efficiency is achieved through high resource utilization (i.e., goodput) and minimal retransmission effort.

2 2

2 2

( ) ( ),

( ) ( )i i

i i

Throughput aFairness ALPHA Fairness

n Throughput n a= =

AIRA: The Decrease Rule

• The Decrease Rule applies, at the end of a Round, when p_i > p_{i-1}.

In order to reduce the gap between the Throughput and Desired Throughput Slopes, the Additive Increase factor should decrease, according to Equation:

1 ∙(1 )i i ia a p+ = −

AIRA: The Decrease Rule

AIRA: The Increase Rule

• The Increase Rule applies, at the end of a Round, when p_i < p_{i-1}.

In order to reduce the gap between the Desired Throughput - Throughput Slopes, the Additive Increase factor should decrease, according to Equation:

1 1∙(1 )i i i ia a p p+ −= + −

AIRA: The Increase Rule

AIRA: Lower Bounds

• We bound a to 0.1 in order to avoid flow starvation (i.e., a = 0). Therefore, a_{greatest reduction} = 0.9.

• Furthermore, we complement this absolute bound with another, dynamically-adjustable lower bound, which depends on the number of completed Rounds and is called Minimum Additive Increase Limit (MAIL). The reason is twofold:

• Avoid drastic responses due to symptomatic events, and

• Prevent short-flow starvation.

AIRA: Lower BoundsMinimum Additive Increase Limit

(MAIL)

where R_N is the number of completed rounds.

• MAIL exhibits one desirable system property: it favors long-RTT flows over short-RTT ones.

1 ∙ %min greatest reduction NMAIL a a R= = −

AIRA: Response to Contention Decrease

• AIRA fails to exploit extra available resources when contention decreases rapidly (i.e., p_{i-1} - p_i is a small value, incapable of accelerating a fast enough).

• We apply a Reset Condition to AIRA, in order to prevent bandwidth wastage in case of contention decrease.

AIRA: Response to Contention Decrease

AIRA Reset Condition.• Assuming that a flow's cwnd oscillates between

W/2 and W before the contention decrease event, the flow will have the opportunity to expand its cwnd to W + W/2 = 3W/2 after 1/3 of the participating flows leave the system. When this happens, AIRA resets a to 1.

• System-wise, we assume that AIRA should exploit bandwidth fastly iff (n-x/n)< 2/3, where n is the total number of participating flows and x is the number of flows who end their task and leave the system. In this case (i.e., when x > n/3), AIRA resets a to 1.

Simulation ResultsScenario Setup

4 FlowsDiverse­RTT (60ms – 220ms)DropTail (RED similar results)Simulation Time: 300s

Results

Long­RTT flows (flows 3­4) operate with higher AI factors:Fairness is higher for AIRA

Goodput

Retr.

AIMD 200 KB/s 2009

AIRA 209 KB/s 943

Simulation Results

Conclusive Remarks

• AIRA improves Stability (through far less retransmissions).

• AIRA improves Efficiency (through less retransmissions and slightly increased Goodput).

• As contention increases, the ALPHA Fairness Index decreases indicating higher fairness for long-RTT flows, against short-RTT flows.

Conclusions

• We have shown that analytical rules can be derived for accelerating, either positively or negatively, the increase rate of AIMD in accordance with network dynamics.

• We found that the "blind" Additive Increase rule can become an obstacle for the performance of TCP, especially when contention increases.

• Instead, sophisticated, contention-aware additive increase rates may preserve system stability and reduce retransmission effort, without reducing the goodput performance of TCP.

Conclusions

We have shown that for varying increase rates within the same system:

1. Fairness and Stability improves when contention increases.

2. Efficiency can be improved through more sophisticated adaptive Additive Increase schemes.