1. Improving VoIP Call Capacity over IEEE 802.11 Networks
Yeonsik Jeong, Sandeep Kakumanu,
Cheng-Lin Tsao, and Raghupathy Sivakumar
GNAN Research Group
Georgia Institute of Technology
The title of the presentation I will give to you is “Improving VoIP Call Capacity over IEEE Networks”.
My name is Yeonsik Jeong, currently a research scientist at Georgia Institute of Technology.

2. Outline Call Capacity Study Solution Basis Algorithms
Simulations/Experimentation/Analysis
Solution Basis
Algorithms
Performance Evaluation
Conclusion
The outline of the presentation is as follows.
At first I will show VoIP call capacity study using simulations, experimentation, and analysis. From the analysis I will derive a motivation to the solutions and propose and describe three algorithms. Through ns-2 simulation, I will show performance improvement obtained by those algorithms. Finally I will conclude with summary.

3. Call Capacity Study: Simulation
Network Conditions
IEEE b network with 11 Mbps link
Codec Spec of G.711
Frame interval: 10 ms → Frame rate: 100
Frame size: 80 bytes
Data rate: 64 Kbps
Expected number of VoIP calls
85 calls
Actual number of VoIP calls in ns-2 simulation
5 calls in ideal condition
At first I will show VoIP call capacity observed over IEEE b network with 11 Mbps link. The voice codec used is G.711, of which frame interval is 10 ms, frame size is 80 bytes, and consequently data rate is 64 Kbps. Under this given condition, the expected number of VoIP calls is 85 calls. However, the actual number of VoIP calls in ns-2 simulation using a simple topology is only 5 calls.

4. Call Capacity Study: Experimentation
Testbed Diagram
To verify the capacity obtained in ns-2 simulation, I built a real testbed consisting of several laptops and desktops, routers, and AP which are all operated on Linux. A real VoIP call is setup between two laptops located in wired domain and wireless domain, respectively.

5. Call Capacity Study: Experimentation
Call Setup
One real VoIP call via Kphone and SIP Express Router
A number of emulated calls via Iperf
Bidirectional CBR/UDP traffic
Frame size: 92 bytes (considering 12 byte RTP header)
Data rate: 73.6 Kbps
Actual number of VoIP calls in testbed experimentation
5 calls
One real VoIP call is setup via KPhone and SIP Express Router and a number of calls are emulated via Iperf. For Iperf configuration, I use bidirectional CBR/UDP traffic with a frame size of 92 bytes considering 12 byte RTP header, whose data rate is 73.6 Kbps.
In this testbed, I measured MOS value whenever one call is added to the network. The graph shows the real capacity observed is only 5 calls, which is the same as the simulation result.

6. Call Capacity Study: Analysis
Terminology (also used as a Metric)
Maximum Frame Rate (MFR)
Captures the frame rate that can be expected at each layer L
Minimum Required Transmission Delay (mRTD)
Gives time taken to transmit the PDU at the corresponding layer
Maximum Number of VoIP Calls (MNVC)
I will briefly show the reason of the low performance. Before the analysis, I will define some terminology, which will be also used as a metric in performance evaluation.
I need to define three terms, which are maximum frame rate, minimum required transmission delay, and maximum number of calls.
Maximum frame rate captures the frame rate that can be expected at each layer L and minimum required transmission delay gives time taken to transmit the PDU at the corresponding layer. These two terms are inverse to each other.
Finally, the most important maximum number of VoIP calls, abbreviated as MNVC, can be defined as MFR divided by 2k, where k is the frame rate.

7. Call Capacity Study: Analysis
Application Layer Capacity
Ideal throughput
The potential MNVC at the application layer is 85 calls
Using those terms, I will show the capacity at each layer in the protocol stack.
First, application layer capacity means ideal throughput. MFR is R over 8D. From the calculation, MNVC of application is 85.

8. Call Capacity Study: Analysis
Impact of Transport and Network Layers
Add protocol headers to the frame
Transport layer and network layer add each protocol header to the frame. From the calculation, MNVC of IP is 57.
The potential MNVC at the network layer, taking into account all the overheads of RTP, UDP, and IP headers, is 57 calls

9. Call Capacity Study: Analysis
Impact of MAC Layer
Adds considerable overhead to the frame including MAC header, MAC backoff time, MAC ACK, and inter-transmission times (DIFS and SIFS)
MAC layer adds considerable overhead to the frame including MAC header, MAC backoff time, MAC ACK, and DIFS/SIFS. From the calculation, MNVC of MAC dramatically drops to 6.
The potential MNVC at the MAC layer, taking into account all the overheads of the higher layers, DIFS, SIFS, backoff delay, and ACK, is 6 calls

10. Call Capacity Study: Analysis
Impact of Physical Layer
Adds long preamble known as PLCP header transmitted at the basic rate (1 Mbps)
Physical layer adds long preamble which is transmitted at the basic rate. From the calculation, MNVC of physical layer is only 5.
We can know that the expected number of VoIP calls is 85 but the actual number is only 5 and main reason is MAC overhead.
The potential MNVC at the PHY layer, taking into account all the overheads of the higher layers and PLCP header, is 5 calls

11. Solution Basis Final Equation for MNVC of Physical Layer
Five Schemes Possible to Improve MNVC
ACK Aggregation (AA): results in the reduction of TACK
Frame Aggregation (FA): decreases k
Link Adaptation (LA): can control R
Time Saving (TS): reduces TDIFS
Header Compression (HC): reduces ∑H(L)
From the analysis, I will derive a motivation to the solutions.
Final equation for MNVC of physical layer is as follows. To increase the MNVC, several parameters within the equation can be controlled, which consequently induces five schemes possible.
First, we can reduce TACK, which is related to ACK aggregation.
Second, we can decrease the frame rate k, which results in frame aggregation.
Third, we can control the data rate R, which is realized through link adaptation.
Fourth, TDIFS can be reduced, which we call time saving.
Fifth, through header compression, we can decrease the total summation of the header length.

12. Solution Basis Impact of Each Scheme on Performance Improvement
Proposed Solutions
ACK Aggregation (AA)
Frame Aggregation (FA)
Link Adaptation (LA)
I will show the impact of each five scheme on MNVC improvement, respectively. The maximum improvement possible is 7.1 in AA, 43.2 in FA, 5.1 in LA, 5.3 in TS, and 5.2 in HC. I will ignore TS and HC, and therefore I will consider the first three schemes for algorithm development and proposed them in the next slides.

13. ACK Aggregation (AA) Motivation and Description
AA refers to sending a single ACK for a block of n frames
Adaptive AA algorithm uses variable block size based on the received block ACK information
Increases the block size upon receiving a block ACK with all successes
Decreases the block size on receiving a block ACK with even a single frame loss
ACK aggregation means sending a single ACK for a block of n frames, which can reduce the time of ACK. Figure shows the throughput vs. delay budget with varying block sizes. The block size is related to delay budget. Smaller block size results in smaller delay but lower throughput and larger block size vice versa.
From the observation that there is a cross-over point, I can motivate an algorithm with adaptive block size. According to the received block ACK information, block size is increased when receiving a block ACK with all successes and decreased when receiving one with even a single frame error.

14. Frame Aggregation (FA)
Motivation and Description
FA refers to fusing multiple frames destined to the same end user into a single large frame
Enhanced piggybacking aggregates frames only when required
To improve the performance of upstream flow, a client maintains a variable which holds the number of aggregated frames in the previously received frame from the AP
Frame aggregation means fusing multiple frames destined to the same end user into a single large frame, which is motivated from the figure. The figure shows the throughput increases as the frame size increases and finally saturates.
Basically there are a lot of frames to be aggregated at AP. However it is not the case in client. To balance the aggregate size between downstream and upstream, each client maintains a variable holding the aggregate size in the previously received frame from the AP. Using this simple technique, I can improve the performance more than simple aggregation algorithm.

15. Link Adaptation (LA) Motivation and Description
LA refers to changing the transmission rate for the data frames
SARF (Size-aware Auto Rate Fallback) algorithm is based on ARF, but it considers channel condition as well as the frame size
If a small frame is in error then there is a high probability of error for a large frame as well, and when a large frame is successful, there is also a high probability of success for a small frame
Link adaptation means changing the transmission rate according to channel condition. In this figure I performed two simulations with different frame sizes. For example, in 6 dB the best rate is 11 Mbps for 92 byte frames but 5.5 Mbps for 1472 byte frames. This can be a motivation for size-aware link adaptation algorithm. The proposed algorithm, SARF, is based on the well-known ARF algorithm, but considers the frame size. Basic concept is as follows. If a small frame is in error then there is a high probability of error for a large frame as well, and when a large frame is successful, there is also a high probability of success for a small frame

16. Performance Evaluation
ns-2 Simulation
VoIP traffic
Bi-directional CBR/UDP traffic with a frame size of 92 bytes
Background traffic
Bi-directional CBR/UDP traffic with a frame size of 1472 bytes
Number of wireless nodes
Single in AA and LA, multiple in FA
Loss rate threshold: 2 %
Metrics
Maximum Frame Rate (MFR)
A fine grained metric for small improvement in AA and LA
Maximum Number of VoIP Calls (MNVC)
A metric for improvement of one call or more in FA
Now I will show the performance evaluation of the proposed three algorithms using ns-2 simulation. For VoIP and background traffic, I use bi-directional CBR/UDP traffic with a frame size of 92 bytes and 1472 bytes, respectively. Loss rate threshold is 2 % and delay budget is varying or 60 ms when fixed.
Two metrics are used. One is MFR and the other is MNVC. MFR is a fine grained metric for small improvement in AA and LA, on the other hand MNVC is used as a metric for improvement of one call or more in FA.

17. Performance Evaluation
ACK Aggregation (AA)
Erroneous channels with BER of 10-5 and 10-4
I performed ACK aggregation simulation under erroneous channels with BER of 10-5 and ACK aggregation shows better performance than original IEEE b. In addition, adaptive AA follows the envelope of two graphs with different fixed block sizes (4 and 8) and gives the best performance.
The Adaptive AA follows the envelope of two graphs with different fixed block sizes and gives the best performance

18. Performance Evaluation
Frame Aggregation (FA)
Erroneous channel with a BER of 10-5
This is the result of frame aggregation with an erroneous channel with a BER of The first simulation is about MNVC vs. delay budget. The FA gives considerable improvement of up to 18 calls for a delay budget of 60 ms. In case of background traffic, it also shows better performance, for example, with 3 Mbps of background traffic, it still gives 7 calls compared to the only 1 calls of the original IEEE b.
The FA gives 18 calls without any background traffic and 7 calls with a background traffic of 3 Mbps for a delay budget of 60 ms

19. Performance Evaluation
Link Adaptation (LA)
Erroneous channel with varying SNR condition
This is the result for the link adaptation under an erroneous channel with varying SNR condition. SARF shows similar performance to ARF in throughput, but better performance in QoS metric such as the number of frames sent at wrong rate.
The SARF shows similar performance to ARF in MFR but better performance in QoS metric as the number of frames sent at wrong rate

20. GNAN Research Group, Georgia Tech
Conclusion
Summary
Analyze the reasons of the inferior performance of VoIP over IEEE networks
Setup an experimental testbed to verify the analysis
Propose three algorithms at the MAC layer and show the performance improvement
Ongoing Work for Software Implementation
Implement the FA into Linux kernel in the real testbed
Show the performance improvement of 17 calls
This is the summary for the presentation. …
I also implemented the FA algorithm in Linux kernel and found the performance improvement up to 17 calls in the real testbed.
Thank you for hearing the presentation.
GNAN Research Group, Georgia Tech

21. Analysis/Simulation with Different Codec
Codec Spec of GSM 6.10
Frame interval: 20 ms → Frame rate: 50
Frame size: 33 bytes
Data rate: 13.2 Kbps
MNVC in b with GSM 6.10
Simulation of FA