1. Delaying Tranmissions in Data Communication Networks to Improve Transport-Layer Performance
Presented by LINGLING MENG( ), XUN XU( )
Based on the previous paper on IEEE journal on selected areas in communications, VOL. 29, NO.5, MAY 2011

2. Objection and Theoretical Support
As we know, the performance of transport layer protocols relies on the performance achieved by packet delivery in the network layer. For reliable data transfer, lost packets require retarnsmissions and thus cause very long delays.
In our work, we discuss how to improve the throughput performance of transport layer protocols by adjusting the operation of the network at the network layer.
We propose to introduce intentional delay in network layer transmissions to reduce the occurrence of traffic bursts, which have detrimental effects on transport layer performance as they can lead to packet loss due to buffer overflow. And we focus on networks with small packet buffers.
We present a novel pacing algorithm that decreases the burstiness of network traffic by delaying packet based on the length of the local packet buffer as we called Queue Length Based Pacing (QLBP).

3. Theoretical Support II
Pacing Network Architecture:
A QLBP system includes a delay queue, a rate controller and three parameters: µmax ,µmin , Qmax
This pacing process can be implemented on the outgoing interfaces of the non-optical routers. This routers have sufficiently large buffers that allow moderate traffic bursts to be absorbed and paced without packet loss to reduce the burstiness of traffics before it enters the small-buffer networks.
The general idea is to dynamically adjust the sending rate of a queue according to the queue length, rather tahn to send packets at a constant rate.

4. Theoretical Support III and Experted Result
Packets arrive at a certain rate on the input link and are stored in the delay queue. The output rate µ(t) is controlled by a rate controller according to the queue length q(t): if 0≤ q(t)≤ Qmax, µ(t) is calculated in a deterministic way; if Qmax ≤ q(t) ≤ Qlim, µ(t) is set to the capacity C of the outgoing link.
µ(t)= [(µmax -µmin )/ Qmax ] *q(t) +µmin , 0≤ q(t)≤ Qmax or µ(t)= C, otherwise.
Now we talk about how to decide the length of delay queue q(t).
When λ(t)< µmin, q(t)=0 and µ(t)= µmin . As a result, no packets are paced and the actual output rate is still λ(t). When λ(t) exceeds µmin, a queue begins to bulid up, which cause µ(t) to increse to follow λ(t). When the edge is reached, µ(t)=λ(t), and the corresponding q(t) is given by q(t)= {[λ(t)- µmin]/ (µmax - µmin )}*Qmax . It's possible for λ(t) to be even larger than C. In this case, q(t) grows up to Qlim .

5. Simulation Conditions
Packet Number: pkts
Packet Size: 1000 Bytes
Input Side: 200 pkts/sec
On_time: 1 sec Off_time: 9 sec
Qmax: 100, 200, … 1000
Processing rate: Max 125 pkts/sec
Min 1.25 pkts/sec

6. Simulation Results (1)
Coefficient of Variation (CV): a normalised measure of dispersion of a probability distribution
Definition taken from Wikipedia

7. Simulation Results (2)
To test the burstiness, the max number of departure packets within 1 second is also measured:

8. Simulation Results (3)
Some bad news:

9. Conclusion
Queue Length Based Pacing (QLBP) helps to smooth the traffic at a cost of introducing delay and longer input buffer queue.
In real traffic, it may bring more benefits, such as less packet loss and retransmission