(IJCSIS) International Journal of Computer Science and Information Security, Vol. 11, No.

12, December 2013

A Cross Layer UDP-IP protocol for Efficient Congestion Control in Wireless Networks
Uma S V, K S Gurumurthy
Department of ECE, University Visvesvaraya College of Engineering, Bangalore University Bangalore, India
1

Uma S V1, K S Gurumurthy2
Department of ECE, R N Shetty Institute of Technology, 2Reva Institute of Technology and Management, Visvesvaraya Technological University, Bangalore, India

Abstract—Unlike static wired networks, mobile wireless networks present a big challenge to congestion and flow control algorithms as wireless links are in a constant competition to access the shared radio medium. The transport layer along with IP layer plays a major role in Congestion control applications in all such networks. In this research, a twofold approach is used for more efficient Congestion Control. First, a Dual bit Congestion Control Protocol (DBCC) that uses two ECN bits in the IP header of a pair of packets as feedback is used. This approach differentiates between the error and congestion-caused losses, and is therefore capable of operating in all wireless environments including encrypted wireless networks. Secondly, for better QoS and fairshare of bandwidth in mobile multimedia wireless networks, a combined mechanism, called the Proportional and Derivative algorithm [PDA] is proposed at the transport layer for UDP traffic congestion control. This approach relies on the buffer occupancy to compute the supported rate by a router on the connection path, carries back this information to the traffic source to adapt its actual transmission rate to the network conditions. The PDA algorithm can be implemented at the transport layer of the base station in order to ensure a fair share of the 802.11 bandwidth between the different UDP-based flows. We demonstrate the performance improvements of the cross layer approach as compared to DPCP and VCP through simulation and also the effectiveness of the combined strategy in reducing Network Congestion. Keywords—congestion; explicit congestion bits [ECN]; transport layer; Internet Protocol [IP]; transmission rate;

I.

INTRODUCTION

Mobile wireless networks present a big challenge to congestion and flow control algorithms as wireless links are in a constant competition to access the shared radio medium and are also affected severely by random losses. Furthermore, CSMA/CA-based wireless links suffer dramatically from neighborhood interferences, where packet transmission decisions are sensibly affected by carrier sensing within the interference range as well as the use of the RTS/CTS mechanism. Besides, the presence of random losses due to the wireless transmission properties is a non-negligible phenomenon that worsens the performances of such networks. All these factors contribute in the well-known performance degradation of wireless wide-spreading networks. Therefore,

congestion control has to be considered in a different manner compared to wired networks, and should be intensively investigated. The issue of Congestion control in wireless networks is often dealt with two prominent techniques. First are the Explicit congestion control schemes, where routers play an important role, since they are well located to react to a congestion state. When congestion occurs, they explicitly inform the end hosts of this state by explicit control messages. Feedback control information can be binary or explicit. One such scheme is the Explicit Congestion Notification (ECN), where each router marks a passing IP packet's header when an incipient congestion is detected. The end hosts react to an ECN-marked packet by reducing their transmission rates. A second approach is derived from ATM forum’s rate-based congestion control algorithms. In these schemes, the routers explicitly determine the permissible throughput of the bottlenecks and assign to each flow its fairshare according to the available bandwidth. In this work, a cross layer approach involving marking the IP packet headers efficiently for congestion notification with differentiation of the type of losses and then using a new algorithm for allotting of fairshare bandwidth among the competing UDP flows is proposed. In the first part of this paper, we propose a new congestion control protocol, Dual bit Congestion Control Protocol (DBCC) with two new schemes: i) A novel distributed scheme that allows for operation within wireless encrypted networks, and ii) A new heuristic loss differentiating scheme that can distinguish between error caused loss and congestion-caused loss. In DBCC, a congestion level is carried by a chain of two packets and each packet provides two bits out of four bits of information associated with a congestion level. The routers compute and distribute a congestion signal into two packets. The congestion level can be specified by concatenating a group of two ECN bits together from a pair of packets at an end node. Incorporated with a novel heuristic algorithm, DBCC can appropriately react to congestion caused loss while avoiding unnecessary reductions of the sending window sizes in response to error-caused loss.

58

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 11, No. 12, December 2013

In the second part of this paper a new router based congestion control algorithm, the Proportional and Derivative algorithm (PDA) is proposed. This PDA monitors the level of network congestion through the occupancy of the buffer which is maintained with the control target. Based on the difference between the present and target occupancy, the PDA controller associated with each link computes periodically a fair rate) and forwards the result to the next gateways till the destination. Using a feedback scheme, the destination supplies the source with the minimal received fairshare. In their turn, sources have to adapt their transmission rate according to received fair rate. We show in this paper through simulation results that the Combined DBCC and PDA algorithms guarantee efficiency and fairness simultaneously reducing the Network congestion effectively. Besides, through analysis, choice of control gain leading to system stability in term of traffic transmission rate has been made. The paper is structured as follows: Section II presents the related work, Section III discusses the proposed cross layer approach for congestion control. In Section IV, we present the implementation and simulation results. Section V concludes the paper. II. RELATED WORK

The two issues addressed in this work have been addressed individually in the past. Over the last few years, abundant techniques have been developed to improve the efficiency and fairness of TCP. A network explicit feedback mechanism based on link throughput measurement was developed in [1]. Another explicit call admission control scheme, called EXACT is proposed in [2]. If the proposed explicit congestion control schemes succeed to reach efficiency and fairness, none of them is dealing with the stability criterion, which is one of the most important issues in highly dynamic wireless networks. Examples include the works of [3], [4], [5] using algorithms to adaptively adjust the sending window size, and [6], [7], [8], [9] employing alternative congestion signals. However, due to their integrated controller design, these techniques often fail to achieve both efficiency and fairness [10]. By decoupling efficiency control from fairness, eXplicit Congestion-control Protocol (XCP) [11] and Variablestructure Congestion-control Protocol (VCP) [12] can achieve high utilization, low persistent queue length, insignificant packet loss rate, and sound fairness depending on the heterogeneity characteristics of a network. While XCP requires the use of a large number of IP packet header bits to relay congestion information thereby introducing significant deployment obstacles, VCP only uses the two existing ECN bits in the IP header to encapsulate three congestion levels hence presenting a more practical alternative of deployment than XCP. However, VCP can only deliver limited feedback to end hosts since two bits can at most represent four levels of congestion. In order to avoid sudden bursts, VCP has to control the growth of transmission rates by setting artificial

bounds. This yields slow convergence speeds and high transition times. Moreover, due to the use of fixed parameters for fairness control, VCP exhibits poor fairness characteristics in high delay networks. Very recently, several works have attempted addressing the problem associated with VCP limitations by increasing the amount of feedback. While the work in MLCP [13] proposes using 3 bits to represent the Load Factor (LF), the UNO framework [14] proposes another alternative to increase the amount of feedback by passively utilizing information in IP Identification (IPID) field. In contrast, DPCP [15] proposes a distributed framework that allows for using no more than 2 ECN-bits to deliver a 4-bit representation of the LF. That said, DPCP needs to access partial information in the TCP header in order to be able to efficiently distribute and reassemble the LF. However, in encrypted networks protected by IPSec, TCP header information is lost when crossing encryption boundaries. Thus, DPCP cannot operate in such encrypted networks. Furthermore, wireless networks are characterized by fading related error-caused loss in addition to queuing related congestion-caused loss. Experiments have shown that the performance of any congestion control protocols relies on appropriate reaction to loss according to its source. Like VCP, DPCP reacts to loss without differentiating between the sources of loss and thus performs inefficiently over wireless networks. Considering the issue of fair share, it has been proven in [16] and [17] that TFRC does not always meet its fair share when the network conditions are dynamic and may present TCP-unfriendliness behavior. In [18], the TFRC performance degradation in wireless environment is highlighted and found to be due to the so-called RTS/CTS congestion induced problem. Previous research in TCP and TFRC performance improvement over wireless networks includes investigating loss discrimination algorithms (LDA) in order to distinguish losses due to congestion from those caused by random wireless errors [19, 20, 21, 22, 23, 24, 25, 26]. Moreover, several other adaptive RTP-based congestion control schemes use a similar approach to react to a loss situation in the network. A first set try to investigate the correlation between the ROTT (Relative One-way Trip Time) and a congestion loss. Extensive experimental results conducted in [19, 20] show that spike-trains observed in a ROTT-graph are only related to congestion losses and not to random losses. Congestion control Schemes like PASTRA [27] and VTP [28] take profit of the ROTT loss discrimination algorithm to find congestion signals. Another approach, called the inter-arrival scheme, uses the time between the arrivals of two consecutive packets as a congestion indication [19, 20]. In [29], Vicente et al. present the design of LDA+, a loss-delay based congestion algorithm, based on the inter-arrival scheme. An improvement of the Datagram Congestion Control Protocol (DCCP) is also considered in [30] showing that the bandwidth utilization is improved by more than 30% and up to 50% in significant setups. The PDA has previously been adapted to ABR flows

59

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 11, No. 12, December 2013

congestion control in ATM networks [31, 32] and proved its efficiency in terms of fairness, accuracy and stability when deployed for congestion control for UDP traffic in wired environments [33]. However, despite their simplicity and transparency, these implicit flow control approaches can be tricky and unreliable over wireless networks, because large delay fluctuations are inherent to such types of networks. Moreover, some of the previously-cited works report inaccuracies in these differentiators [34]. III. THE PROPOSED CROSS LAYER DESIGN The proposed cross layer approach to congestion control in wireless networks can be realized in two stages: Stage I: A Dual bit Congestion Control Protocol (DBCC) that uses two ECN bits in the IP header of a pair of packets as feedback is used. Stage II: Secondly, for better QoS and fairshare of bandwidth in mobile multimedia wireless networks, a combined mechanism, called the Proportional and Derivative algorithm [PDA] is proposed at the transport layer for UDP traffic congestion control.
This combined approach to dealing with congestion at both the IP and the UDP layers also takes into account only the congestion caused packet losses and helps in mitigating congestion in all wireless networks including encrypted ones very effectively. Each stage is implemented individually and explained below along with simulation results and then the unified technique is implemented and the results of the combined technique are discussed in the end.

Figure 1: Architecture of the Initial approach

Figure 2: Architecture of the Cross Layer Approach

A. The Dual Bit Congestion Control[DBCC] The design of DBCC is motivated by two observations. First, most feedback based congestion control protocols either require the use of multiple bits in the IP header or even access to headers of the protocols above the IP layer, thereby facing deployment challenges in encrypted networks. Second, most congestion control protocols are designed for wired networks and treat both types of loss as congestion caused loss. While error-caused losses are typically absent in wired networks, they are common in wireless networks. Experiments show that reacting to error-caused and congestion caused loss, can significantly decrease the performance of any congestion control protocol. Thus, the target operating environments of DBCC are IP-based wireless networks including encrypted wireless networks. This means that only eight bits of the IP header, two ECN bits and six Type of Service (ToS) bits, can bypass the encryption boundaries and are available for end to end signaling. As the ToS bits are reserved for signaling differentiated services as opposed to congestion control, DBCC will only use the two ECN bits of the IP packet header for carrying congestion control signaling feedback. Overview: Relying on two new schemes, DBCC works efficiently in all wireless networks. 1. First and albeit the fact that DBCC uses a double packet four bit representation of the LF, it introduces a packet ordering management strategy that is quite distinct. It only utilizes the information available in the IP header and only manipulates two existing ECN bits to carry congestion information. The IPID field of the IP header originating from a host is either monotonically increasing or chosen uniformly at random. In either case, the LSB of IPID flips over quickly enough to be used for signaling MSP/LSP. Specifically, DBCC only uses the LSB of the IPID field. Further, the use of IPID field bits is passive, i.e., the bit values are inspected but not changed by DBCC. A packet with an LSB value of zero is used as the MSP and a packet with an LSB value of one is used as the LSP. If the IPID is increased incrementally, the LSB bit flips over for any pair of consecutive packets which is perfect for differentiating MSP from LSP. If it is varied randomly, then DBCC uses the first packet with an LSB value of zero for carrying MSP and the first packet with an LSB value of one for carrying LSP. As evidenced in our experiments, it is safe to assume that bit flips, with a probability of 0.5, occur quickly enough with respect to necessary congestion reaction speed especially over large BDP networks. 2. Second, DBCC utilizes a heuristic scheme for differentiating error-caused loss from congestion-caused loss. This heuristic scheme runs at the transmitting side and maintains the history information of congestion status over the bottleneck link of a path. Upon detection of loss, the heuristic scheme makes an identification of the source of loss based on the saved history information. Given the fact that the feedback is updated with the receipt of every ACK, it is reasonable to

60

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 11, No. 12, December 2013

assume that the congestion status of a network can be continuously tracked by the sender. It is especially important to realize that a congestion caused loss event has a much longer duration than an error caused loss event. Relying on the above fact, the heuristic algorithm of DBCC assumes that a sender can identify the cause of a loss by keeping track of the status of the network. In order to track the status of the network, the heuristic algorithm proposes maintaining a revolving congestion history Bit Map (BM) of size N at the sending side. Upon the receipt of an ACK, the bit at position BM(1) is dropped, the bit at position BM(i) with i ∈ {1,…,N} is shifted to the left so it takes the position of bit BM(i − 1), and the bit at position BM(N) is set to 1 if the new ACK indicates congestion or otherwise to 0. If at any time, the right most T consecutive bits with T ≤ N are set to 1 in the bit map, a binary flag called Congestion Flag (CF) is set to 1. Otherwise, the flag is set to 0. Upon detection of a loss, if CF flag is set, then the loss is safely determined as a congestion-caused loss triggering a multiplicative decrease operation to cwnd(congestion window). Otherwise, the loss is considered to be an error-caused loss and the sender simply maintains the current cwnd. In the case of DBCC, the link LF is encapsulated in ACK packets and the OVER LOAD represents a LF beyond 100%. Thus, OVER LOAD is used as the indicator of congestion. According to our experiments, setting N to 32 and T to 16 represent optimal choices. We note that with these choices of values, maintaining the revolving history bit map only requires 4 bytes of storage on a per flow basis. While N should essentially be a function of flow cwnd, we set the value of N to 32 for the convenience of implementation. We also note that the value of cwnd for larger flows could be easily scaled to fit the 32 bits of N. Fig. 3 illustrates the operation of the heuristic algorithm of DBCC.

Figure 3: Illustration of the loss differentiating heuristic algorithm in DBCC

3. Finally, the security mode operation IPSec operates in two modes: transport mode and tunnel mode. In the transport mode, the original IP header is kept after getting authenticated by IPSec. Thus, DBCC can still access IPID and ECN bits as usual in IPSec transport mode. In contrast, the entire packet is encrypted and authenticated in IPSec tunnel mode. As a result, the original IP header becomes invisible in the encrypted packet. Since the LSB bit of the IPID in the original IP header may not necessarily be the same as the one in the new IP header, DBCC utilizes the IPID only on the Cipher Text (CT) side but not on the Plain Text (PT) side for packet ordering. As DBCC will be installed and configured at the IPSec router, it is safe to assume that DBCC will have access to both CT

and PT headers of a packet. Specifically, DBCC provides two router modules: i) Security Module (SM) running only on IPSec routers that cooperates with IPSec gateways, and ii) Normal Module (NM) running on both IPSec gateways and other routers. Assuming an FTP or a comparable connection has been established, the flow of events at the IPSec gateways is as follows: i) A DBCC packet arrives at the ingress of an IPSec gateway. Before the packet goes to the IPSec module for encryption, DBCC SM will first catch the packet, save the packet ordering information, i.e., MSP/LSP and the value of the LF as indicated in the ECN bits. Then DBCC SM delivers the packet to the IPSec module. After the new IP header is generated and ready to be transmitted through the tunnel, DBCC SM catches the outgoing packet again and encodes ECN bits with MSB/LSB bits of the saved LF depending on the LSB bit of the IPID in the new IP header. Note that, after the original IP header is encrypted, DBCC has no idea of whether the new packet is a TCP packet or a packet using another protocol, e.g., UDP. Thus, DBCC encodes ECN bits regardless of the original protocol type, which introduces overhead for non-TCP packets. In fact, this is the tradeoff between efficiency and protocol complexity. That said, we note that the resulting overhead is not significant because i) it is only introduced when transmitting over IPSec tunnels; and ii) it is only associated with the operations of encoding an LF. ii) At the output interface of the ingress IPSec gateway, DBCC NM takes over. DBCC NM compares the LF in the packet with the LF of its downstream link interface and updates the LF in the packet if necessary iii) At the intermediate router on the CT side, DBCC NM operates as DPCP router module except that DBCC uses the LSB bit of IPID to identify MSP/LSP. iv) At the egress of the IPSec gateway and before the encrypted packet goes to the IPSec module for decryption, DBCC SM will catch the packet and save the LF value as indicated by the ECN bits of the packet. Note that after the packet is decrypted, the IPSec module will copy the ECN bits from the new IP header to the original IP header on the PT side. However, the packet ordering information cannot be simply transferred to the PT side. While DBCC SM can access both CT and PT side, DBCC SM dedicates to change the contents of the packet as minimally as possible. Simply put, DBCC SM does not directly pass any bits from the CT side to the PT side. Note that, the LSB bit of the IPID in the original IP header is not necessarily the same as the one in the new IP header. Thus, instead of changing the value of the LSB bit of the IPID field in the original IP header for the purpose of matching the one in the IP header used by the IPSec tunnel, DBCC uses the relative order of the TCP seq and ack numbers as the indication of MSP/LSP after the original IP header is retrieved. In this way, DBCC will not change any bits in the IP header of the decrypted packet. Furthermore, DBCC SM has to keep a copy of the LF of the upstream link of the egress

61

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 11, No. 12, December 2013

IPSec gateway for each IPSec tunnel. DBCC SM inspects the ECN bits in the packet and compares it with the MSP/LSP of the saved copy of the LF of its upstream link. Based on the results of the comparison, DBCC SM manipulates the seq and ack numbers in order to mark the packet as MSP or LSP. Then the packet is delivered to DBCC NM. DBCC NM updates the ECN bits according to the LF of its downstream link following the operating mechanism of DPCP. B. THE PDA CONTROLLER FOR UDP TRAFFIC This controller helps achieve better QoS and fairshare of bandwidth in mobile multimedia wireless networks. This PDA uses quite a standard approach: the level of network congestion is monitored through the occupancy x of the buffer which is maintained with the control target being xo. Based on the difference between x and xo, the PDA controller associated with each link computes periodically at time n a fair rate q(n) and forwards the result to the next gateways till the destination. Using a feedback scheme, a destination supplies the source with the minimal received q(n). In their turn, sources have to adapt their transmission rate according to received fair rate. 1) Buffer equations Shown in Fig. 4, each node has a congestion controller associated to its outgoing link i, this controller calculates at each control period n a supported fair rate qi(n) based on local information: the difference between the buffer occupancy xi(n) and a fixed threshold xo, as well as the control decision at present and in the finite past: qi(n-1), qi(n-2),…… qi(n-k) Then, the dynamics of buffer i is described by the following equation: ‫ݍ‬௜ ሺ݊ + 1ሻ = ܵܽ‫ݐ‬௤బ ቐ‫ݍ‬௜ ሺ݊ሻ − ෍ ∝௝ ‫ݔ‬௜ ሺ௡ି௝ሻି௫ − ෍ ߚ௞ ‫ݍ‬௜ ሺ݊ − ݇ሻቑ , ݅ ∈ ܰ Where j and k are non negative integers.
௝ୀ଴ ௃ ௝ୀ଴ ௃
బሻ

0 ݂݅ ‫ < ݖ‬0 ܵܽ‫ݐ‬௔ ሺ‫ݖ‬ሻ = ൝ ܽ ݂݅ ‫ܽ > ݖ‬ ‫ݐ݋ ݖ‬ℎ݁‫݁ݏ݅ݓݎ‬ The saturation function is introduced to impose bounds on the computed qi(n): the lower bound zero keeps qi(n) positive, whereas the upper bound qO limits the sending rate of connections with non-congested paths. As stipulated in [19], in order to ensure the system stability, the coefficients αj and βk must satisfy the following conditions: ∑௃ ௝ୀ଴ ߙ௝ > 0, ∑௄ ௞ୀ଴ ߚ௞ = 0 ሺ2ሻ

The first order derivative PDA controller is for the case of j=0, k=0, and is so governed by the following equation: ‫ݍ‬௜ ሺ݊ + 1ሻ = ܵܽ‫ݐ‬௤బ ሼሺ‫ݍ‬௜ ሺ݊ሻ −∝଴ ሺ‫ݔ‬௜ ሺ݊ሻ − ‫ ݔ‬଴ ሻ − ߚ଴ ‫ݍ‬௜ ሺ݊ሻሽ ሺ3ሻ ‫ݍ‬௜ ሺ݊ + 1ሻ − ܵܽ‫ݐ‬௤బ ሼሺ‫ݍ‬௜ ሺ݊ሻ −∝଴ ሺ‫ݔ‬௜ ሺ݊ሻ − ‫ ݔ‬଴ ሻሽ ሺ4ሻ

And, according to conditions stipulated by equation (2), equation (3) leads to:

With the above design, the system does not match the expected stability criteria: first, it exhibits an instable behavior with several burst losses .Second, the buffer occupancy x does not oscillate in the neighboring of the threshold xu. These are confirmed by experimental results, which now demonstrate the need to introduce a second derivative component in the PRDR controller equation as shown below:
௢ ሼሺ‫ݍ‬ሺ݊ሻ ‫ݍ‬ሺ݊ + 1ሻ = ‫ݐܽݏ‬௤ −∝଴ ሺ‫ݔ‬ሺ݊ሻ − ‫ ݔ‬଴ ሻ − ߙூ ሺ‫ݔ‬ሺ݊ − ‫ܫ‬ሻ − ‫ ݔ‬଴ ሻሽ

‫ݔ‬ሺ݊ + 1ሻ = ‫ݔ‬ሺ݊ሻ + ‫ݍ‬ሺ݊ + 1ሻ − ߤሺ݊ሻ

ሺ5ሻ

ሺ1ሻ

where ∝଴ and ߙூ , are the two first derivative control gains and ߤሺ݊ሻ denotes the rate at which new connections have been admitted to the network during the time interval [n, n+1]. Solving the equations for proper stability (i.e poles of the polynomial have negative real roots), the control gains have to obey the following conditions: −2 < ߙூ < 0 ܽ݊݀ ∝଴ +∝ூ > 0 ܽ݊݀ ∝଴ <∝ூ + 4

ሺ6ሻ

The control gains ∝଴ and ∝ூ , values can be selected among the set of values defined in the domain D presented in Fig.5.An appropriate choice of ∝଴ and ∝ூ , will be obtained by simulations in the next section.
Figure 4. A PDA controller.

The saturation function is such that:

62

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 11, No. 12, December 2013

Figure 5. Stability domain of the control gains ∝଴ and ∝ூ

2) Adaptation of PDA for UDP flow control In the transport layer, the Real-Time Protocol (RTP) is used since it can be implemented on top of UDP/IP stack. For a feedback mechanism, we chose to use the RTCP protocol, since it doesn’t consume an additive bandwidth and doesn't inject too much supplementary traffic in the network. Indeed, for a sender transmission rate of 2Mbps, RTCP sender reports sent every 0.1 seconds add only a supplementary traffic of 2.5%. The design of the PDA algorithm is as follows: an incoming traffic source ‘s’ connected to a destination ‘d’ expresses its initial desired rate Rd in a field of an RTCP "Application-specific" control packet which is forwarded to next router until it reaches the destination d. Each intermediate node m on the path from s to d captures the value of Rd carried in the RTCP control packet and substitutes it to the locally computed fair share rate qm if smaller and forwards this information to its neighbour node m+I. Finally, the control packet reaches the destination d with the smallest value of qm on the connection path. Then, the destination sends the received fair share back to source s in an RTCP RR packet, which replaces its actual transmission rate Rd by the received qm.. Different sources periodically send RTCP control packets every TControl. The choice of the TControl value affects sensibly both the transient response (settling time and initial connection parameters) and the control overhead due to the computation and the transmission of the feedback information. Faster updates periods lead to shorter settling time, more rapid steady state and smaller buffer overshoot, whereas a smaller TControl value increases the control overhead. IV. PERFORMANCE EVALUATION IN WIRELESS ENVIRONMENT In this section, simulation studies and experimental studies of DBCC and PDA are presented, first individually and then combined. A. DBCC DBCC is implemented in both NS-2 simulator and Linux Kernel. Performance of DBCC, DPCP, and VCP are compared in terms of efficiency and fairness. Since DBCC is proposed as an extension of DPCP for encrypted wireless networks, our target environment is characterized by moderate bandwidth (2 − 10Mbps), low delay (200 − 1000ms) lossy links. The wireless effects are introduced by utilizing the temporally correlated Gilbert Elliott (GE) model [35]. We now compare the performance of DPCP and VCP over a four bottleneck parking lot topology as illustrated by Fig. 6(a).

All of the links have a one-way delay of 250ms and a bandwidth of 4Mbps except L2 that has a bandwidth of 2Mbps. The GE model is applied on a per link basis in order to introduce an average loss rate 5% for each link. There are two types of aggregate FTP flows traversing the topology. The first type is referred to as a Long Flow and represents the combined traffic of 30 FTP flows traversing all of the links in the forward left-to-right direction. The second type is referred as to as a Local Flow. There are four Local Flows each of which representing 10 FTP flows traversing each individual link in the forward direction. Except those flows that traverse link L2 and start after 1000 seconds, all other Local Flows start at the beginning of the experiments.

Figure 6. An illustration of a) Parking lot

b)Dumbbell topologies

Note that if no wireless loss is introduced, DBCC and DPCP achieve nearly identical performance as they share same control policy. With the heuristic scheme, DBCC can significantly improve the performance of DPCP over a lossy link. Fig. 7 shows the bandwidth split ratio of VCP, DPCP, and DBCC respectively. Ideally, during the first 1000 seconds, both Long and Local Flows are to equally split the bandwidth of a shared link. Starting from 1000-th second when an extra Local Flow starts at link #2, the utilization of Long Flows at Link #0 should drop to 25% while the utilization of Local Flows should go up to 75%.

Figure. 7. a) A performance comparison of DBCC, DPCP, and VCP over link #0. 7. b) A performance comparison of DBCC, DPCP, and VCP over link #2.

63

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 11, No. 12, December 2013

In Fig.7a, VCP exhibits a biased fairness characteristic splitting the bandwidth of link #0 with a ratio of 15 to 1. While DPCP demonstrates a significantly better fairness characteristic than VCP, it shows inefficiency in terms of the bandwidth utilization due to the effect of its reaction to loss. In contrast, DBCC shows both good fairness and efficiency. At link #2, we expect to see a near 100% bandwidth utilization for Long Flows during the first 1000 seconds and a split of 50% in the last 1000 seconds between Long and Local Flow when the Local Flow joins. As illustrated by Fig. 7b, both DBCC and DPCP show good fairness and responsiveness, although DBCC outperforms DPCP in terms of bandwidth utilization. To the contrary, the bandwidth split ratio does not change even when Local Flows are turned on in the case of VCP showing that VCP fails to achieve fairness in high BDP multiple bottleneck topologies serving flows with heterogeneous RTTs. 1) Experimental Studies In this subsection, we describe our implementation of DBCC in the Linux Kernel. The implementation approach follows that of VCP as described in [36]. Again, we introduce packet loss using our GE error model implementation in the Linux Kernel. In this section, we do present our experimental study conducted over a real testbed comparing the performance of VCP, DPCP, and DBCC. We present the results associated with a single bottleneck scenario. We use a dumbbell topology (Fig. 7.b) with the settings used for experiments matching those of [36]. Though not shown here, the performance of DBCC in multi-bottleneck scenarios follows the same pattern shown in our simulation studies. Fig. 8 compares the bandwidth utilization of VCP, DPCP, and DBCC over the single bottleneck link. In our experiments, a loss rate of up to 30% is introduced. Thus, both DPCP and VCP fail to open the cwnd efficiently in the absence of the heuristic scheme, and therefore exhibit a low utilization characteristic. Note that while DPCP achieves higher bandwidth utilization than VCP, it demonstrates oscillations due to its inappropriate reaction to error-caused loss. The improvement comes from the faster recovery speed of DPCP in contrast to VCP. In contrast, DBCC can identify the source of a loss and ignore error-caused loss. In the figure, DBCC can achieve significantly better bandwidth utilization than both DPCP and VCP although it shows oscillations due to the associated retransmissions and timeouts. It is clearly seen from the results that DBCC overcomes the limitations of DPCP by using an alternative packet ordering management scheme. Rather than accessing the TCP header, DBCC passively inspected the LSB bit of the IPID field in the IP packet header to identify whether a packet is the MSP or LSP in a packet pair sequence. Furthermore, DBCC utilized a heuristic loss identification scheme to differentiate error-type and appropriately react to loss.

Figure 8. A performance comparison of DBCC, DPCP, and VCP over the bottleneck link of our experimental dumbbell topology.

We implemented DBCC in both NS-2 and the Linux Kernel. Through simulation we demonstrated that the fairness and efficiency characteristics of DBCC are comparable to those of DPCP in wired networks. We also demonstrated that in high BDP networks, both DBCC and DPCP significantly outperform VCP in terms of fairness and efficiency. As the main differentiating factors, we showed that i) unlike DPCP, DBCC can operate over IPSec encrypted networks, and ii) relying on its heuristic loss identification algorithm, DBCC can significantly outperform DPCP in wireless environments characterized by tandem loss B. PDA Network topology and test configurations In order to study the performances of the PDA algorithm in a wireless environment, we considered in our simulations a heterogeneous topology with base station (BS) on the NS-2.31 network simulator that implements the topology described in Fig. 9.Here, N sources located in the wired side of the network initiate k CBR/RTP traffic to N wireless destinations. All sources are connected to the BS via a gateway G with 15Mbps and 10 ms-delayed links. The channel bandwidth is 11 Mbps and the payload size of each data packet is 1500 bytes. All results are given from five times simulations with 300 seconds duration each. The DSDV routing protocol is used. The BS implements the second-order PDA, depicted in equation (5) and (6): for a queue length of 200 packets, the threshold on queue occupancy ‫ݔ‬଴ is set to 180 packets. Since the maximal good-throughput of 11 Mbps wireless channels is about 4.5 Mbps, we judged that a value of 4.5 *0.9 is a good choice for the target rate. The control period T control is set to 100 ms and the control gain (∝଴ to 0.8). In our simulations, the following several important performance metrics are evaluated: • Good end-to-end throughput ‫ݎ‬௜ ; the amount of data delivered to the destination for flow ݅ ሺ‫ܰ ≤ ݅ ≤ ܫ‬ሻ • BS buffer occupancy x • Stability measured as the standard deviation for ‫ݎ‬௜ , series, denoted ߪ௜ • Fairness index: we use the Jain's fairness index used in [37] and defined by:

64

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 11, No. 12, December 2013

݂=

ଶ ሺ∑ே ௜ୀଵ ‫ݎ‬௜ ሻ ଶ ܰ ∑ே ௜ୀଵ ‫ݎ‬௜

Figure 11 Instantaneous buffer occupancy in presence of four competing RTP flows Figure 9. A PDA controller

1) Channel allocation Initially consider the network described in Fig.9 where the BS uses a classical drop-tail queuing algorithm. Fig.10 plots the instantaneous rates of 4 competing RTP flows, and Fig.11 plots the instantaneous BS buffer occupancy, starting at 0, 20s, 30s and 40s respectively. It is clear from Fig. 10 and 11 that the wireless channel is unfairly allocated between the four flows since it represents the bottleneck of the network. By analyzing the NS-2 trace file, it was found that the major cause of packet drops is buffer overflow (IFQ). When the drop-tail queuing discipline is used, traffic gets synchronized, which allows the first and the fourth flow to monopolize the queue space and consequently get the maximal channel bandwidth allocation. The mean value of the fairness index samples is 0.8. Moreover, the BS buffer is saturated all the time (x = 200 packets). The unfairness problem is also revealed for the ten competing RTP flows depicted in Fig. 12. As we can see, eight flows are overwhelming the bandwidth (getting almost 500 kbps of bandwidth), whereas the two remaining flows are roughly discriminated (obtaining only 200 kbps). In Fig. 13, we replaced the fourth UDP flow by a single TCP connection. We can notice that the TCP flow is totally starved since it is unable to send any packets (TCP rate is zero). The fairness index is only 0.6. In addition, all the flows see frequent burst losses (occurring at times 160s, 170s and 260s), which leads to a simultaneous decrease of the present flow rates, confirming the hypothesis of the global synchronization problem related to the drop-tail queuing policy.

2) Channel allocation with II Order PDA algorithm We present here simulation results with four competing RTP flows, using the same network configuration for different values of control gains ∝଴ and ∝ூ , within the domain of Figure 6. The fair rate for the four RTP flows is then: ܴி = 4.5‫ × ݏ݌ܾܯ‬0.9 = 1.102‫ݏ݌ܾܯ‬ 4

Figure 12. Instantaneous rates in presence of ten competing RTP flows

Figure 13. Instantaneous rates in presence of three competing RTP flows and a single TCP connection

The Table I below provides the performance criteria described in Fig. 10 for different values of ∝଴ and ∝ூ . For all the cases, the measured fairness index f is very close to 1(0.998).

Figure 10. Instantaneous rates in the presence of 4 competing RTP flows

65

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 11, No. 12, December 2013

TABLE I Standard Deviations for different values of ∝଴ and ∝ூ

C. Combined Simulation of DBCC and PDA Finally both DBCC and the PDA were simultaneously implemented in the wired cum wireless environment as shown in Fig.9. The BW Utilization/rate factor significantly improves when both are implemented together even in highly congested bottlenecks. As demonstrated in Fig.16 the congestion was completely eliminated 98% of the time along with the different flows getting fair allocation of BW. The fairshare allotted to the 4 different flows is calculated here also and it was observed that bandwidth utilization is most efficient at certain specific allotted rates and accordingly the simulations were conducted for an allotted rate of 5.25Mbps, the new rate of 1.246Mbps was allotted to all the flows. ܴி = 5.25‫ × ݏ݌ܾܯ‬0.95 = 1.247‫ݏ݌ܾܯ‬ 4

As an illustration, Figures 14 and 15 plot the instantaneous allocated rate for the RTP flows respectively for cases 2 an 7. From the plots, we can conclude that setting ∝଴ ‫ ݋ݐ‬2.5 and ∝ூ ‫ ݋ݐ‬− 0.5 is the best choice since the system has the best behavior in terms of stability (not oscillatory) and convergence. Moreover, the buffer occupancy is better controlled concerning the fixed threshold xo of 180 packets.

The buffer occupancy for different types of RTP flows is as shown. Results show that our mechanism can fairly allocate wireless bandwidth resource in heterogeneous networks and converges to a steady state whenever the input traffic parameters change.
R 1400 a 1200 t e 1000 800 k b 600 p 400 s 200 0 0 50 100 150 200 250 300 Time (s) 250 200 150 Buffer status 100 50 0 0 50 100 150 200 250 300 Time (s)

RTP-1
RTP-2

RTP-3 RTP-4

Figure 14. Instantaneous RTP rates or ∝଴ ‫ ݋ݐ‬2 and ∝ூ ‫ ݋ݐ‬− 1

Figure 15. Instantaneous RTP rates for ∝଴ ‫ ݋ݐ‬2.5 and ∝ூ ‫ ݋ݐ‬− 0.5

Figure 16. Instantaneous RTP rates and buffer occupancy for ∝଴ ‫ ݋ݐ‬2.5 and ∝ூ ‫ ݋ݐ‬− 0.5

66

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 11, No. 12, December 2013

V.

CONCLUSIONS

Mobile wireless networks present a big challenge to congestion and flow control algorithms as these links are in a constant competition to access the shared radio medium. In this research, a twofold approach using combined Dual bit Congestion Control [DBCC] at the IP layer and Proportional and derivative algorithm [PDA] at the Transport layer is used for more efficient Congestion Control. First, DBCC involving two ECN bits in the IP header of a pair of packets is used for congestion situation feedback. This approach differentiates between the error and congestion-caused losses, and is therefore capable of operating in all wireless environments including encrypted wireless networks. Secondly, for better QoS and fairshare of bandwidth in mobile multimedia networks, the PDA mechanism is proposed at the transport layer for UDP traffic congestion control. Simulation results have shown the efficiency of both techniques individually in comparison with other standard existing techniques and also that of the combined technique where both are implemented together. There is a rate improvement of about 5.7% as compared to the previous implementation of the individual techniques and Congestion is avoided 98% of the time. REFERENCES
1. Manthos Kazantzidis, Mario Gerla, "End-to-end versus Explicit Feedback Measurement in 802.11 Networks", in Proceedings of the Seventh International Symposium on Computers and Communications, Italy, July 2002. Kai Chen, Klara Nahrstedt, Nitin Vaidya, "The Utility of Explicit RateBased Flow Control in Mobile Ad Hoc Networks", in Proceeding of IEEE Wireless Communications and Networking Conference, Atlanta, USA, March 2004. L. Xu, K. Harfoush, and I. Rhee, “Binary Increase Congestion Control (BIC) for Fast Long-Distance Networks,” in Proc. of the IEEE INFOCOM, 2004. I. Rhee and L. Xu, “CUBIC: A New TCP-Friendly High-Speed TCP Variant,” in Proc. of the PFLDNet’05, Feb. 2005. S. Floyd, “High Speed TCP for Large Congestion Windows,” Aug. 2002. D. Leith and R. Shorten, “H-TCP: TCP for High-speed and Longdistance Networks,” in Proc. of the PFLDNet’04, Feb. 2004. T. Kelly, “Scalable TCP: Improving Performance in HighSpeed Wide Area Networks,” Feb. 2003, available at http://wwwlce.eng.cam.ac.uk/ctk21/scalable/. C. Jin, D. Wei, and S. Low, “FAST TCP: Motivation, Architecture, Algorithms, Performance,” in Proc. of IEEE INFOCOM, 2004. S. Bhandarkar, S. Jain, and A. Reddy, “Improving TCP Performance in High Bandwidth High RTT Links Using Layered Congestion Control,” in Proc. of the FLDNet’05, Feb. 2005. M. Goutelle, Y. Gu, and E. He, “A Survey of Transport Protocols other than Standard TCP,” 2004, https://forge.gridforum.org/forum/forum.php?forum id=410. D. Katabi, M. Handley, and C. Rohrs, “Congestion Control for High Bandwidth-Delay Product Networks,” in Proc. ACM SIGCOMM, Aug.2002. Y. Xia, L. Subramanian, I. Stoica, and S. Kalyanaraman, “One More Bit Is Enough,” in Proc. ACM SIGCOMM, 2005, Aug. 2005 I. A. Qazi and T. Znati, “On the design of load factor based congestion control protocols for next-generation networks,” in Proc. of the IEEE INFOCOM 2008, Apr. 2008.

2.

3.

4. 5. 6. 7.

8. 9.

10.

11.

12. 13.

14. N. Vasic, S. Kuntimaddi, and D. Kostic, “One Bit Is Enough: a Framework for Deploying Explicit Feedback Congestion Control Protocols,” in Proc. of the First International Conference on COMmunication Systems and NETworkS (COMSNETS), Jan. 2009. 15. X. Li and H. Yousefi’zadeh, “Distributed ECN-Based Congestion Control,” in Proc. of the IEEE ICC 2009, June 2009. 16. Bansal, H. Balakrishnan, S. Floyd, and S. Shenker, "Dynamic behavior of slowly-responsive congestion control algorithms," in Proceedings of ACM SIGCOMM, San Diego, California, USA, August 2001. 17. Y. Yang, M. Kim, and S. S. Lam, "Transient behaviors of tcp friendly congestion control protocols," in Proceedings of IEEE INFOCOM, Anchorage, Alaska, USA, April 2001. 18. M. Li, E. Agu, M. Claypool , R. Kinicki, "Performance Enhancement of TFRC in Wireless Networks", Worcester Polytechnic Institute Technical Report, December 2003. 19. S. Cen, P. Cosman, , G. Voelker, "End-to-end Differentiation of Congestion and Wireless Losses", in IEEE/ACM Transactions on Networking, vol. II, issue 5,2003. 20. C. Parsa and 1. Garcia-Luna-Aceves, "Differentiating congestion vs. random loss: A method for improving TCP performance over wireless links" in IEEE WCNC, Chicago, September 2000. 21. J. Liu, I. Matta, and M. Crovella, "End-to-end inference of loss nature in a hybrid wired/wireless environment," in Proceedings of WiOpt, France, March 2003. 22. Pierre Geurts, Ibtissam EI Khayat, Guy Leduc, "A machine learning approach to improve congestion control over wireless computer networks", In Proceedings of IEEE International Conference on Data Mining, Brighton, UK, November 2004. 23. Bin Zhou, Cheng Peng Fu Li, V.O.K. , "TFRC Veno: An Enhancement of TCP Friendly Rate Control over Wired/Wireless Networks", in the Proceedings of the IEEE International Conference on Network Protocols, Bejing, China, October 2007. 24. Bin Zhou Cheng Peng Fu Chiew Tong Lau Chuan Ileng Foh, "An Enhancement of TFRC over Wireless Networks", in Proceedings of the Wireless Communications and Networking Conference, Hong-Kong, March 2007. 25. S. Pack, X. Shen, 1. W. Mark, and L. Cai, "A Two-Phase Loss Differentiation Algorithm for Improving TFRC Performance in IEEE 802.11 WLANs," in IEEE Transactions on Wireless Communications, February 2007. 26. Neng-Chung Wang , Jong-Shin Chen , Yung-Fa Huang , Chi-Lun Chiou, "Performance Enhancement of TCP in Dynamic Bandwidth Wired and Wireless Networks", in Wireless Personal Communications: An International Journal archive, volume 47, issue 3, November 2008. 27. Y. Tobe, Y. Tamura, A. Molano, S. Ghosh, H. Tokuda, "Achieving Moderate Fairness for UDP flows by Path-Status Classification", in the Proceedings of the 25th Annual IEEE Conference on Digital Local Computer Networks, florida, USA, November 2000. 28. G. Yang, M.Gerla, M.Y; Sanadidi, "T. Smooth and efficient Real-Time Video Transport in the presence of Wireless errors", in ACM Transactions on Multimedia computing communications and Applications (TOMCCAP), vol 2, issue 2, May 2006 29. Vicente E. Mujica V., Dorgham Sisalem, Radu Popescu-Zeletin, Adam Wolisz, TCP-Friendly Congestion Control over Wireless Networks, in Proceedings of European Wireless, Barcelona, Spain, February 2004. 30. Ijaz Haider Naqvi , Tanguy Perennou , "A DCCP Congestion Control Mechanism for Wired- cum-Wireless Environments", in the Proceeding of the IEEE Wireless Communications and Networking Conference, Hong-Kong, March 2007. 31. I. Lengliz, F. Kamoun, "A rate-based flow control method for ABR service in ATM networks", in Computer Networks Journal, Vol. 34, No.1, pp. 129-138, July 2000. 32. I. Lengliz, F. Kamoun, "System Stability with the PD ABR Flow Control Algorithm", in the Proceedings of the Fifth IEEE Symposium on Computer and Communications, ISCC, France, July 2000. 33. I. Lengliz, Abir Ben Ali, F. Kamoun, "A novel issue for multimedia traffic control in the internet", In Technical Program of Soft'Com, Italy, October 2004.

67

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

(IJCSIS) International Journal of Computer Science and Information Security, Vol. 11, No. 12, December 2013

34. Kamal Deep Singh, David Ros, Laurent Toutain , Cesar Viho, "Improvement of Multimedia Streaming using Estimation of Wireless losses", IRISA Research report, March 2006. 35. X. Li and H. Yousefi’zadeh, “An Implementation and Experimental Study of the Variable-Structure Congestion Control Protocol (VCP),” in Proc. Of the IEEE MILCOM, 2007, Oct. 2007. 36. H. Yousefi’zadeh, X. Li, and A. Habibi, “An End-to-End Cross-Layer Profiling Study of Congestion Control in High BDP Wireless Networks,” in Proc. of the IEEE WCNC, 2007, Mar. 2007. 37. Ling-Jyh Chen; Chih-Wei Sung; Hao-Hsiang Hung; Sun, T. Cheng-Fu Chou, "TSProbe: A Link Capacity Estimation Tool for Time-Slotted Wireless Networks", in Proceedings of IFIP International Conference on Wireless and Optical Communications Networks, Singapore, July 2007. Author’s Profiles

Uma Satyanarayana Visweswaraiya is currently pursuing her research on “Congestion Control Techniques in Communication Networks” under Dr. K S Gurumurthy in Bangalore University. She is also working as an Associate Professor in the Eectronics and Communication Department of RNSIT, Bangalore since 2006, teaching both UG and PG students in core subjects like Analog and Digital electronics, A & D Communication, Computer Communication Networks and CMOSRF Circuit Design. She has published two books, the most recent one being “Constraint Based design of Communication Networks using GA”, Lambert Academic Publishing, Germany, 2012, and three papers in International Journals like Springer-Verlag and IJSER. Her research interests include Communication Networks and Signal Processing. Gurumurthy Satyanarayana Rao Kargal has completed his B.E degree in E & CE, from MCE, HASSAN, Mysore University, and M.E degree from IIT, ROORKEE and a PhD from IISc, Bangalore-12. His experience is as Administrator/Coordinator/Specialist and Professor in ECE at UVCE, BU, Bangalore, INDIA .He was also heading the department. In addition to teaching and guiding PhD/ME/BE students he was responsible for the smooth running of the department. He has over 40 international publications to his credit. Presently he is a Professor in the ECE department of Reva Institute of Technology and Management. His specialization is VLSI Design and Communication Networks.

68

http://sites.google.com/site/ijcsis/ ISSN 1947-5500

Sign up to vote on this title
UsefulNot useful

Master Your Semester with Scribd & The New York Times

Special offer for students: Only $4.99/month.

Master Your Semester with a Special Offer from Scribd & The New York Times

Cancel anytime.