Throughput Enhancement of Simultaneous Multipath Communication Using Modified Fast

Retransmit Scheme Samiullah Khan\ Syed Zubair Ahmed2, Mohammad Abdul Qadi?

Centre for Distributed and Semantic Computing, Department of Computer Science,

Mohammad Ali Jinnah University, Islamabad, Pakistan

lsamikhan_1982@yahoo.com; 2szubair@jinnah.edu.pk; 3aqadir@jinnah.edu.pk

Abstract- The increased availability of communication resource through simultaneous use of multiple network interfaces of mobile and fIXed devices has enabled higher bandwidth for streaming applications. The out-of-sequence arrival on such multipath transport configurations remains the major chaUenge to be overcome before the available potential is maximized. In this paper, we propose Modified Fast Retransmit (MFR) Scheme for the multi-homed stream

control transmission protocol (SCTP) that minimizes the impact of possible out-of-sequence (OOS) arrival at the destination. The proposed scheme isolates local and global OOS arrivals and adapts fast retransmission strategy for global OOS arrival, in accordance to the end-to-end communication path characteristics. The essence of higher availability of bandwidth is also signified through the modified congestion control mechanism. The proposed scheme produces large boost in throughput of the streams transported through multiple paths, as compared to other schemes; such as Concurrent Multipath Transmission.

Kqwords----Concurrent multipath transport; modified fast retransmit; congestion control; out-of-Sequence arrival; throughput maximization

I. INTRODUCTION

The Stream Control Transmission Protocol (SCTP); as standardized by Internet Engineering Task Force (IETF) provides the multi-homing feature modern devices to use one interface as primary path while others may be used as alternate backup paths [1]. The Protocol Engineering laboratory (PEL), University of Delaware has proposed an extension of SCTP; named as Concurrent Multipath Transfer (CMT) [2] [3]. The CMT is capable of the simultaneous transmission of data on multiple paths using available multiple interfaces. The bandwidth aggregation capability of CMT has been considered as a suitable choice in host of application scenarios; such as, high performance video streaming, video conferencing; IPTV and others. The same approach is also effective for mobile sessions where multi-path transport can significantly absorb handover latencies by redirecting flow density to other alternate paths during mobility events [4].

One of the major challenges in multipath transport of flows is the OOS arrival of segments at the receiver side [8]. The OOS arrival in such scenarios, can be both on individual paths as well as unified receive window, where segments wait for in-order delivery to the applications.

978-1-61284-941-6/111$26.00 ©2011 IEEE

9

The conventional fast retransmit strategy used in SCTP can cause drastic retransmission due to the varying network conditions and hence not only reduce effective throughput for the session, but also wastes useful communication resource. A large number of segment waiting for in-order delivery may also reduce the buffering capacity of the receiver and may cause very low values of advertised receive window. These events may eventually lead to significantly degraded growth of sender's congestion window (cwnd) and subsequently reduces throughput. Iyengar et al. have overcome the unnecessary fast retransmission problem along with providing alternate retransmission strategies [2]. In spite of the fact that cwnd doesn't collapse seriously; the throughput from combined bandwidth of multiple paths remain much lower than its available capacity.

In this paper, we have proposed a MFR scheme to control the retransmission process of congestion control by further refining the duplicate acknowledgement (DUP_ACK) handling procedure and the response of the sender. The global OOS; caused by the varying E2E delay of multiple paths has been handled separately to accommodate variations of E2E path characteristics. The cwnd has also been handled on the basis of aggregated bandwidth of all the links to allow it more aggressive expansion, in case receiver has advertised sufficient capacity. The rest of the paper is organized as follows. Section II discusses some state-of-the-art schemes proposed till date for concurrent transport of streams. Section III describes the proposed scheme. The simulation results of the proposed MFR scheme are compared and analysed with the other schemes in Section IV. We conclude in Section V.

II. RELATED WORK

The SCTP with its multi-homing and multi-streaming features has been actively studied as a high bandwidth transport service in recent years. The initial multi-homed support was restricted to a single interface use at a time with other interfaces providing standby support to the primary path [6]. The CMT as proposed by Iyengar et al. [2]; was one concrete solution that modified standard SCTP for simultaneous use of multiple paths through available interfaces. There are three basic issues in the basic CMT proposal. These include; premature fast retransmission, Overlay conservative cwnd growth and

Increased traffic volume due to fewer delayed acknowledgements. The authors proposed 8pilt Fast Retransmit (8FR) algorithm to overcome the first problem. The second problem was catered through special Congestion Update for CMT (CUC) messages. Finally; for the third problem Delayed Acknowledgement for CMT (DAC) was proposed. In addition to these, further five algorithms were proposed for ensuring comprehensive retransmission policies for CMT.

The performance gain in terms of average throughput was studied through simulated in [5]. This study proposed the usefulness of the CMT with limited bandwidth networks with some heuristics. The WCMT-8CTP was proposed to solve the received buffer blocking problem [6]. In WCMT-8CTP, certain data streams were bound to specific paths and used multiple active paths to transfer data simultaneously with limited received buffer space. This specific configuration produced acceptable results in ad-hoc wireless networks but throughput suffered in high capacity networks. We have used unlimited receiver buffer space to focus our study on network behaviour.

In [8], concurrent multi-path 8CTP (cmp8CTP) has been proposed to improve the data transfer rate. The proposed scheme introduced several advanced structures such as multiple-buffer for sub-streams, multi-state management, two-level sequence numbers and a cooperative 8ACK strategy to boast bandwidth aggregation. However, all these approaches have their own merits and demerits. The re-sequencing delay issue at the receiver has been discussed in [9]. The re­sequencing delay has been modelled through mean delay to reduce reordering complexity.

III. MFR 8CHEME

In this section, we present proposed MFR scheme for performance of CMT scheme in presence of 008 arrival of segments due to different path characteristics. The 008 arrival of segments may be further categorised into local and global 008. Let there be segments with sequence number 1 to 10 that are scheduled to be transmitted through path A and segments with sequence number 11 to 12 are scheduled to be transmitted to the receiver using path B. The received 008 segments at the receiver can be classified in two types. In case, the segments arrive in the range from 1-10, not in sequence, they are classified as local 008 arrival. Fig.l shows the same scenario where packets 4, 6, 7 and 9 are local 008 arrivals.

�. Time 1 2 3 4 5 6 7 8 9

Fig. I Local OOS

In case, the segment numbers are in the range of 1 to 12 (not completely in their incremental order) the arrived segments are termed as a set of global 008 segments. Fig. 2 shows the same scenario where segment 4, 11 and 8, 12 constitute a global 008 subsets. In fact we can

10

specify all 008 sub-streams as either local or global in a subset form rather than treating them as a whole.

.�. Time 1 2 3 4 5 6 7 8 9 10 11 12

Fig. 2 Global OOS

A. Motivation/or MFR Scheme

In light of the above classification, The MFR scheme is based upon two observations. 1) The transport layer is responsible for the

transmission of in sequence data segments to the application layer. The transport layer holds all received segments irrespective of their order in an anticipation of slightly delayed arrival of missing segments. In order to minimize extra delay, and buffering cost the missing packets may be retransmitted by some alternate strategy, other than timeout of the segments.

2) The global 008 arrival in multipath communication does not indicate any kind of congestion or loss of segment condition in the network. Therefore, duplicate acknowledgement (dup_ack) may not be needed for these 008 arrivals.

In a reliable transport protocol; like TCP and 8CTP, dup_ack(s) are sent every time an 008 segment arrives. The sender, on receiving these dup_ack(s) decide there retransmission strategies. One such approach is named as fast retransmit strategy that dictates retransmission of the segment whose three dup_acks are received. The underlying assumption in such retransmission is that the segment may have been lost due to congestion as the later three segments have already reached, as indicated by three consecutive dup_acks. This approach saves some time for retransmission as we need to wait for the time out event for retransmission which keeps on chan"iing during the life time of a stream. One major consequence of such approach is the readjustment of cwnd of the stream that may be slashed down according to different proposed algorithms. The same is pictorially presented in Fig. 3. In Fig. 3(a) the fast retransmit is followed by the fast recovery event in which the cwnd is set to half of previous cwnd value. In contrast, Fig. 3(b) depicts modification in fast retransmission process under certain known conditions.

Now considering the scenario of a multi-path transport, the local 008 arrival depicts the same scenario as given in the previous Para, but the case of global 008 does not indicate any congestion event and transmitting dup _ ack in such cases may trigger unnecessary retransmission of segments causing wastage of useful bandwidth if not contributing to congestion. Therefore, the fast recovery event may be modified or avoided to stop unnecessary reduction in cwnd size. The main motivation of the proposed MFR scheme is essentially focused to deal with above mention scenarios. The possible vitiations of E2E path characteristics may lead to diverse spectrum of global 008 arrival. Hence,

distinguished handling of global OOS forms the core of the proposed scheme.

Begin fast Rftraosmit

(Count =3) Begin �Iodified fast

Retransmit (COUDt =3)

Retransmit lost packet

a) b)

Fig. 3 Flow diagram of a) Fast Retransmit b) Modified Fast Retransmit(MFR)

B. Modified Fast Retransmit (MFR)

The unified handling of local and global OSS cause

added delay and result in CMT throughput degradation.

This delay can vary depending upon the bandwidth and

end-to-end (e2e) parameters of the multiple paths.

According to our proposed MFR scheme, the SFR

algorithm is modified in such a way that the missing

packet should be retransmitted with minimum delay. Figure 4 shows the Split Fast Retransmission (SFR)

algorithms with proposed modifications.

Algorithm: Spilt Fast Algorithm (Modified Version)

Input: On receipt of a SACK containing gap reports lSender Side behaviour):

1) For all destination address di • initialize di . saw_newack=F alse;

2) For each TSN ta being acked that has not been acked in any SACK thus far do

Let da be the destination to which ta was sent; Set da.saw_newack = TRUE;

3) For all destinations dn. set dn .hightesUn_sack_for_dest to highest TSN being newly acked on dn.

4) To determine whether missing report count for a TSN tm should be incremented for fast retransmit or should be incremented for Modified Fast

Let dm be the destination to which tm was sent;

If (dm. saw_newack == TRUE) and

(dm .hightesUn_sack_for_dest > tm ) then

Increment missing report of tm 's Fast Retransmit counter.

5) Else

If (dm• saw_newack == False) and (dn .hightesUn_sackJor_dest > tm ) then

Increment missing report for MFR counter

MFR= Fast retransmit count+1;

Fig. 4 Modified Fast Retransmit Algorithm (Modified SFR)

In the step 1-4 [2] of the Fig. 4, the SFR tries to avoid the abnormal trigger of fast retransmission due to

11

Global OOS. We have added step 5 in this algorithm to

minimize the extra delay. We have introduced an MFR parameter that holds state about global OOS arrival. If

the new segment received has higher TSN than any TSN

received on any path then this packet should increment

the MFR counter of missing segment. The incrementing of the MFR counter may have impact of the fast

retransmission counter. If the counter for fast

retransmission of packet is incremented up to 2 then with each arrival of segment will increment the MFR counter

to 3. This help in quick retransmission of the missing

segment.

IV. RESULTS

In this section, we present the experimental setup, topological layout, simulation parameters and results for proposed scheme under simultaneous multi-path transport. The simulation was carried out in well-known network simulator-ns-2 [17]. The simulation model of simultaneous multipath communication was configured between the sender-receivers using two interfaces. The Sender A has two interfaces Al and A2, that communicates with the interface B I and B2 of the receiver B; respectively. The path between the interface Al and BI is named as "Path A" and the path between the Interface A2 and B2 is named "Path B". The independent path was acquired in this case for the simplicity of the experiment. Fig. 5 shows the simulation topology. As show in the Fig 5, Path A has the bandwidth of 0.2 Mbps with 45 ms delays while the Path B has the bandwidth of I Mbps with 45 ms delay. The delay is kept same for the both paths, to reduce the impact of e2e delay variations. This also helps in finding out other factors affecting the throughput of the communicating nodes.

The path B having high bandwidth faces the arrival of OOS packets due to the less bandwidth of path A. The cwnd of path B trigger the abnormal fast retransmission due to OOS packets. This problem was efficiently solved by the CMT as shown in the Fig. 6(a-b). The CMT ignores the effect of global OOS segments toward activation of fast retransmission. This causes the CMT cwnd to show continuity without degradation in the bandwidth in Fig. 6b. In Fig. 6a, the MFR cwnd was less then CMT on path A. In Fig. 6b, the MFR cwnd exceeded the CMT cwnd up to the end of the simulation. If we compare the Fig. 6a with the Fig.7a then this is visible that by enhancing the cwnd of CMT does not help in increasing the throughput of the CMT.

Bandwidth = 0.2 Mb s PI'OP delay = 45ms

Bandwidth = 1 Mb)5 Prop deh,y = 45ms

Fig. 5: CMT network Topology configuration. (Reproduce from [2]).

In Fig.7b, the MFR shows high throughput as compare to the CMT throughput. The throughput degradation in CMT is due to the arrival of global OOS segments. The CMT has overcome the cwnd by reducing it to half on arrival of global OOS segments and by waiting for the arrival of the missing segment. This mechanism results in decrease of overall throughput of CMT. In contrast, MFR scheme tries to identity the missing segment as soon as possible to retransmit it quickly. This approach enhances the MFR throughput significantly under given set of conditions. The throughput is more than thrice of the CMT's throughput, as show in the Fig.7b.

The MFR throughputs have been enhanced at the cost of retransmission of lost segment without implementing fast recovery method. This event prevents the abnormal degradation of MFR's cwnd, which further improve the cwnd and throughput of MFR.

.. III >. e � 0

"0 I:

� I: 0 .� III .. I: 0 u

Path A: cwnd of Expected(SCTP), CMT & MFR

7.00E+0��.w��::===�rlIAT==-=-=-=-=M�CR

/" 6.00E+04

5.00E+04

4.00E+04

3.00E+04

2.00E+04 It'"'

I /

/ �

�_J--------------------1.00E+04 ..1/'

O.OOE+OO

o Time Wee) 15

Fig. 6a: Behaviour of cwnd of Path A in comparison with other schemes.

� " >. e :I 0 ." <: :§ <:

.Q t; " .. <: 0 u

Path B: cwnd of Expected(SCTP), CMT & MFR -- Expected(SCTP) -- CMT -- MFR

7.00E+04

6.00E+04 (

5.00E+04

4.00E+04

I I

3.00E+04

2.00E+04

1.00E+04

O.OOE+OO fJ.

o 10 15 Time (sec)

Fig. 6b: cwnd of MFR on Path B, as compared with other schemes

V. CONCLUSION

The simultaneous multipath communication is one of the emerging transport scenarios with significant potential of

performance boast, particularly in resource scare

networks. In this paper, we have introduced two distinct interpretations of multi-path transport, according to the

arrival of out-of-sequence segments belonging to the

same path or dissimilar path such as Local OOS and

Global OOS respectively. We have proposed the MFR scheme as solutions to the CMT's throughput

degradation. The MFR improves the throughput up to 211

% as compare to the CMT's throughputs. We have also

12

studied the reason for the throughput decline in CMT.

The Transport layer transfers in-sequence segments to the Application layer. However, the SFR triggers the fast

Retransmit only on the Local OOS segments and has to

wait extra time for the reception of the global OOS

segment. This degrades the CMT's throughput.

v III

� '" :.. � Q.

.r; .. " e �

Path A: Througput of Expected(SCTP), CMT & MFR

30000

25000

20000

15000

10000

5000

o

-- Excected(SCTP) -- CMT

I ,

(( """' ............... ...............

10 Time (sec)

...............

Fig. 7a: Path A throughput analysis.

- - - - M FR

............... .....

15

Path B: Througput of Expected(SCTP), CMT & MFR

1.40E+OS -- Expected(SCTP) -- CMT -- MFR

u 5: ..... III � '5 c.

.<: .. " e .<: ...

1.20E+OS l.OOE+OS 8.00E+04 6.00E+04 4.00E+04

It. \ \ \.-,.

2.00E+04 O.OOE+OO

o

-

�"r-"

10 Time (sec)

15

Fig. 7b: Path B Throughput analysis

REFERENCES

[I] R. Stewart, Stream Control Transmission Protocol, RFC 4960 (Proposed Standard), Sep. 2007.

[2] J. R. Iyengar, P. D. Amer, and R. Stewart, Concurrent multi path transfer using SCTP multihoming over independent end-to-end paths, IEEFJACM Trans. Netw., vol. 14, no. 5, pp. 951-964, 2006.

[3] P. Natarajan, J. Iyengar, P. Amer, and R. Stewart, Concurrent Multipath Transfer Using Transport Layer Multi homi ng: Performance Under Network Failures, MILCOM, October 2006.

[4] Syed Zubair Ahmad et al.; Analysis of Multi-Server Scheduling Paradigm for Service Guarantees during Network Mobility, Wireless Personal Communications, DOl: 1O.1007/sI1277-01O-0114-5, Online First.

[5] C.Xin, F.Xuan,"A Simulation-based Study on Performance Profit of Concurrent Multi-path Transmission Schemes", ICISE, 2009.

[6] B. Wang, W. Feng, S.-D. Zhang, H-K. Zhang. Concurrent multi path transfer protocol used in ad hoc networks, lET comm, 2010, Vol.4, Iss.7, pp. 884-893.

[7] B. Wang, W. Feng, S.-D. Zhang, H-K. Zhang." Concurrent multi path transfer protocol used in ad hoc networks" lET comm, 20 I 0,VoI.4,lss. 7,pp. 884-893.

[8] J.Liao, J. Wang, XZhu "cJl1)SCTP: An Extension of SCTP to Suppcrt Concurrent Multi-Path Transfer Communications" 2008. ICC '08. IEEE International Conference, on May 2008 Page( s ):5762 - 5766.

[9] K.Zheng, X.Jiao, M.Liu, Z.Li, An analysis of resequencing delay of reliable transmission protocols over multi path, IEEE communication society, (2010).

[10] "The network simulator - ns-2," http://www.isi.edulnsnamlnsl.