Staggered-TCP for parallel split-sessions across multiple proxies in heterogeneous networks

Saima Zafar, Sana Jabbar Department of Electrical Engineering

National University of Computer & Emerging Sciences, FAST-NU, Lahore, Pakistan

Al-Khawarzmi Institute of Computer Sciences, KICS University of Engineering & Technology, UET

Lahore, Pakistan e-mail: saima_zafar@yahoo.com,

sanajabbar_05@yahoo.com

Ali Hammad Akbar, Noor M. Sheikh Department of Computer Science

University of Engineering & Technology, UET Lahore, Pakistan

Department of Electrical Engineering University of Engineering & Technology, UET

Lahore, Pakistan e-mail: ahakbar@gmail.com,

omanchair@uet.edu.pk

Abstract— Wired-cum-wireless networks are interconnected through proxy or gateway that acts as router and also caters for link MTU mismatch between the two networks. In IPv6 based networks, TCP is mandatory for bulk data transfer from wired to wireless host. This results in end-to-end TCP session through the default proxy. The single proxy supporting a large number of TCP sessions; is vulnerable to buffer overflow which results in end-to-end retransmissions, that degrades TCP performance. The use of split-TCP sessions across proxy has been proposed in order to improve TCP performance in heterogeneous networks by researchers. We assert that the availability of multiple proxies for interconnectivity (that is usually the case) can be exploited in order to make TCP efficient. We propose Staggered-TCP; an architecture that manages parallel split-TCP sessions across multiple proxies. Through simulations we show that for bulk data transfer, our proposed solution optimizes communication, simultaneously providing load balancing services.

Keywords-buffer overflow; multiple-proxies; parallel

sessions; ; latency;

I. INTRODUCTION As a result of burgeoning development and deployment

of wireless networks there is a need to ensure their efficient connectivity with the wired networks. In such heterogeneous networks, both networks communicate within the network as well as with the extraneous networks. The two networks are interconnected through proxy or gateway that acts as a router and also implements fragmentation and reassembly to cater for link MTU mismatch. In IPv6 based networks, TCP is mandatory for reliable data transfer. This results in TCP session through the default proxy. The single proxy supporting a large number of TCP sessions is vulnerable to buffer overflow resulting in end-to-end retransmissions that in turn degrades TCP performance.

Although splitting TCP connections in wired and wireless networks has been proposed for TCP performance improvement, the single proxy remains to be the performance limiting factor. A proxy, usually a layer-five device, is under-utilized in its role as a router only. Also a large number of proxies (which are usually available) can be used to improve TCP efficiency. Our proposed solution is

parallel data transfer across split-TCP connections through a number of proxies. We present a session-layer mechanism that manages split-TCP sessions through multiple proxies and stripes data across them, preserving TCP semantics. Data striping across parallel paths has been suggested in multi-homed hosts for effective bandwidth utilization. However, data striping across a number of paths passing through a number of proxies has never been explored.

We present Staggered-TCP; a session layer mechanism that staggers a single Sender-Receiver (S-R) session into multiple Sender-Proxy-Receiver (S-P-R) split-sessions, across a number of incumbent proxies and strips data across these sessions. Staggered-TCP maintains the semantics of TCP which guarantees flow control, congestion control and reliability. To the best of our knowledge, no prior work has explored the effectiveness of data striping across multiple split-TCP sessions through multiple proxies.

We organize the remainder of the paper as follows. In section 2, related work is discussed. In section 3, we present multiple proxies architecture and discuss the proposed mechanism. In section 4, experimental evaluation of Staggered-TCP mechanism using ns2 simulations is presented. Finally; we summarize our work and conclude the paper in section 5.

II. RELATED WORK Related work comprises of TCP performance

improvement approaches for heterogeneous networks with mismatched characteristics and a comparison of various data striping schemes for multi-homed end hosts.

TCP is said to perform scantily when implemented in hosts present in wired-cum-wireless networks. In such networks, splitting TCP connection into two parts, wired and wireless, improves throughput and fairness. In [1] some of these split TCP approaches have been compared. I-TCP [2-3], Split TCP [4] and Semi-split TCP [5] are some of the variations of this approach. All these schemes improve throughput by replacing end-to-end retransmissions with local retransmissions or retransmissions from proxy to receiver. The performance gain is limited by congestion at the proxy that supports a large number of TCP sessions and

978-1-4244-6949-9/10/$26.00 ©2010 IEEE

also due to asymmetry between links. The proxy can become the bottleneck and buffer overflow can occur at proxy [6] [7]. Efforts have been made in order to make TCP viable and efficient for such networks. In [8], A. Dunkels et al propose Distributed TCP caching as a solution to this problem for heterogeneous networks.

Multiple proxies in heterogeneous networks have been proposed for load balancing as well as for routing and higher degree of reliability [9]. Proxies announce their presence to hosts which select one proxy through which communication takes place. In case that proxy becomes unavailable or its performance degrades, the host selects another proxy out of the advertised proxies. Thus in multiple proxy schemes, one of the proxies has to be selected at a time for communication.

In multi-homed end hosts with multiple interfaces, data striping is proposed for bandwidth aggregation. [10-13] present different data striping schemes across parallel paths between multi-homed sender and receiver. Striping can be implemented at different layers based on application requirements. Striping at higher layers lead to less head-of-line blocking however, it increases application complexity. A. Habib et al [10] suggest session layer striping but do not provide a framework for implementing it. pTCP[11] and mTCP[12] are transport layer striping protocols. pTCP presents a complete framework for transport layer striping and is defined as a wrapper that manages the operation of underlying paths while TCP-v is a TCP-like connection on each path. This requires extensive changes at the transport layer. Session layer striping is suitable for our work because it results in preserving TCP semantics. When FAST-TCP is implemented for multi-homed hosts, some issues have to be addressed and these are discussed in [13]. However, the focus of our work is not TCP implementation in multi-homed hosts but hosts that can be connected through multiple proxies.

III. MODEL AND MECHANISM We present the design of Staggered-TCP; a mechanism

for ingress data flow from wired to wireless network through a number of proxies. Key design elements are as follows.

• Multiple Proxies: We propose the use of multiple proxies in parallel. The role of proxy is enhanced from merely being a router to a device that operates at the session layer and plays its role to avoid buffer overflow, mitigate packet loss and out of order packet delivery. A number of proxies manage TCP sessions discretely in wired and wireless networks in order to make TCP efficient.

• Dynamic rbuf Assignment at the Receiver: The link quality of wireless TCP sessions from proxies to receiver is variable therefore we establish a relationship between the Link Quality Indicator (LQI) and per session receiver buffer. Buffer space for all sessions would be managed by receiver in such a manner that buffer for good quality links would be increased and for bad quality links would be decreased hence less data would be sent on bad quality paths.

• Flow Control: There is a need to reflect buffer constraints of receiver to sender. The flow control is implemented independently in wired and wireless networks. We propose that each proxy instead of advertising its receive window to sender, advertises receiver’s receive window to sender. This way end-to-end flow control would be implemented.

• Congestion control: Each TCP path can have different bandwidth and delay characteristics. If one global congestion window is used, in case of packet loss on a single path, global congestion window would be reduced resulting in unnecessary reduction in throughput on good paths. We suggest independent congestion control for all paths.

• Difference between multi-homing and multiple proxies: Multi-homed hosts have multiple interfaces through which they communicate in parallel, mostly through distinct networks [15]. Parallel data transfer in multi-homed hosts achieves high throughput by utilizing multiple available networks. Our proposed concept of parallel data transfer through multiple proxies is different; we propose data striping across diverse paths through proxies for end hosts which are not essentially multi-homed.

A. Proposed Network Model and Assumptions Figure 1 shows the proposed multiple-paths model connecting the sender (S) in Internet to the receiver (R) in wireless network through a number of proxies (P). The links from S to Ps are wired and may contain multiple intermediate routers and links from Ps to R are wireless, usually pass through multiple hops. Our main interest is ingress traffic from S to R through multiple Ps; this traffic is bulk in nature. We make following assumptions:

• S and R are not essentially multi-homed. • The S, Ps and R all support Staggered-TCP

mechanism. • Ps and R are buffer constrained. • The hosts support Neighbor Discovery Protocols. • Packet size in wired network is much larger than

packet size in wireless network.

Figure 1. Multiple-proxy network model.

B. Staggered-TCP Architecture Figure 2 shows Staggered-TCP architecture, inclusion of

a session layer in the protocol stack to support split- TCP paths across multiple proxies. There are two modules namely Sessions Manager (SM) and TCP Manager (TM). SM maintains a single send buffer and a single receive buffer. When application layer has data to send, the application data is copied onto the send buffer of SM. For one socket opened by an application, SM opens and maintains a number of TM sessions. The status of all TMs is maintained at SM. Each TM opens a TCP socket with the transport layer. SM divides application data into ‘n’ number of data chunks passed to each TM following weighted round robin scheduling. TM implements the functionality of each session which SM opens with the TCP. At the receiver, transport layer data is received by each TM to which it is addressed. SM assembles data chunks into application data before delivering data to the application layer. Each TM processes ACKs independently.

1. SM-TM Interface As shown in Figure 3, we define interface between SM

and each TM by six functions. These functions are call( ), release( ), opened( ), closed( ), read( ), and write( ). The function call( ) is used by SM to open a TM session and release( ) is used to close a TM session. When a complete split-TCP session is established, TM reaches OPENED state. Similarly when a complete split-TCP session is released, TM reaches CLOSED state. When TM state machine reaches the OPENED and CLOSED states respectively, TM informs SM using opened( ) and closed( ) interfaces. Upon receiving the OPENED event from a TM, SM copies the striped data to TM send buffer using write( ). TM then appends its header to this data and passes it to the transport layer. At the receiver, SM fetches data from TM into receive buffer using read( ).

2. Header format The header format for Staggered-TCP is shown in Figure

4. There are flags for connection establishment segments.

Figure 2. Staggered-TCP atchitecture.

Figure 3. SM-TM interface.

Other fields are Intermediate destination port #, Final destination port #, 32-bit SEQ #, 32-bit ACK #, Intermediate destination address and Final destination address. The SEQ # and ACK # are used for in-order data delivery at the receiver. Intermediate destination address and Intermediate destination port # are used for setting up S to P TCP sessions and Final destination address and Final destination port # are used for setting up P to R TCP sessions.

3. Connection management S requests R for data transfer through multiple gateways.

If R turns down the request, S establishes TCP connection with R through the default proxy and Staggered-TCP is not used. If R agrees for parallel data transfer, data transfer follows handshake between S and R. Figure 5 shows timing diagram for connection establishment and data transfer in Staggered-TCP. When information about proxies is available at S, S initiates first TM connection; it opens TCP connection with P1 and sends TM1 ST-SYN to P1 which then opens TCP connection with R and sends TM1 ST-ACK to S. P1 must wait for Wait-State ( ) timeout period to ensure that the “TCP ACK” gets through to R before it sends ST-ACK to S. At this time, first TM connection from S to R through P1 is complete and data transfer begins. One complete TM session comprises of two TCP connections; one in wired network between S and P and other in wireless network between P and R.

Figure 4. Staggered-TCP header format.

Figure 5. Timing diagram for connection establishment and parallel data transfer using Staggered-TCP.

TM connections through subsequent Ps are completed in a similar manner. Data transfer from S to R takes place through Ps till data transfer is complete and connections are released sequentially.

When SM module at sender receives information about proxies, SM creates Staggered-TCP socket with a TCB including P1 IP address, source port #, destination port # and creates first TM TCB by issuing call( ) to it. TM appends Staggered-TCP header to SYN packet which is sent to P1 through TCP socket which it opens with transport layer. On receiving this SYN packet, the TM module in P1 creates TM TCB and returns SYNACK to S, which returns ACK. The TM is in OPEN WAIT state at this time. After TCP connection in wired network is complete from S to P, TM in P performs three way hand shake with R to establish wireless TCP connection from P to R. At this time, ST-ACK is sent back from P to S. TM at S reaches OPENED state and data starts flowing from S to P. SM at S opens subsequent TM sessions through all available proxies. When an application decides to close Staggered-TCP connection, the SM module closes all the TM sessions by issuing release( ) to all TMs sequentially. In a similar manner, each connection closes using TCP closing handshake. When all TMs are closed, SM enters the CLOSED state and informs closed connection to the application.

C. Dynamic Buffer Assignment In Staggered-TPC we establish a relationship between

link quality indicator (LQI) and receiver buffer. The receiver buffer is dynamically adjusted based on channel conditions. Some of the split-TCP sessions are upgraded and some are downgraded in terms of path quality. The receiver buffer is not wasted by reserving a lot of buffer space for bad quality

split-sessions. The mechanism works as follows. Separate equal receive buffer is reserved by receiver for all split-TCP sessions at connection setup time. Later on LQI is measured for each split-TCP session. For those sessions where received data is less, buffer space is reduced and for those where received data is more, buffer space is increased. The receiver buffer is dynamically assigned which facilitates efficient consumption of receiver buffer.

D. Intelligent Data Striping At the session layer, SM module stripes data received by

the application layer into its send buffer. It stripes data and assigns it to each TM module in a round-robin manner. SM does so by inferring the quality of each split-session on the arrival of ST ACK (data), which is the estimation of round-trip-time by SM for each split-session. This way SM keeps a record of each split-session’s quality and stripes data accordingly.

E. Flow Control We implement end-to-end flow control in Staggered-TCP

instead of independent flow control in the two parts of a single split-TCP session. In order to mitigate out-of-order data delivery at R, wireless network constraints should be reflected to S in wired network. Proxies have an important role in implementing this flow control. A proxy on receiving RcvWindow (receive window advertised by R), advertises the same RcvWindow to S. This way, buffer constraints of R are reflected back at S, which adjusts its sending rate accordingly.

RcvWindow is advertised to S in terms of wired link MSS by the proxy. The session layer at proxy translates RcvWindow advertised by R in terms of wired link MSS and advertises this RcvWindow to S. If RcvWindow advertised by SN is in terms of B bytes MSS for wireless network, P translates it in terms of Ethernet MSS which is 1296 bytes. This is accomplished as follows. The RcvWindow advertised by R = x * B bytes. The RcvWindow advertised by P to S is y * 1296 bytes. If these receive windows have to be the same then y * 1296 bytes = x * B bytes which means y = (x * B) / 1296 bytes would be RcvWindow advertised to S by P. Flow control is implemented in this manner for all split-TCP sessions from sender to receiver through all proxies.

F. Congestion Control We support independent congestion control for each TCP

path. A single congestion window for all paths can result in reducing the aggregate throughput even lesser than throughput of a single path. This can happen if one of the paths experiences severe congestion and reduces the single global congestion window although other paths can offer high throughput. This would result in underutilized multiple paths which fails the basic advantage of multiple parallel paths. Instead of a single global congestion window, we recommend independent congestion control on all paths. Each split TCP-session implements its congestion window and each of these are overseen by SM, which stripes data accordingly.

IV. EXPERIMENTAL RESULTS In order to validate benefits rendered by Staggered-TCP

we performed simulations in ns2. We implemented network model shown in Figure 1 and simulated file transfer from sender S to receiver R through gateways P1 and P2 such that file is striped at S. We observed the impact of various parameters on latency of file transfer for single split-TCP session through a proxy and double split-TCP sessions through two proxies. The latency of file transfer considerably reduces with two proxies.

A. Latency of File Transfer when Receiver Buffer is varied Figure 6 and 7 show graphs for latency of file transfer

versus receiver buffer size for file sizes 10 Kbytes, 20 Kbytes and 30 Kbytes respectively. In performance evaluation, receiver buffer is an important parameter because wireless host is usually buffer constrained. We kept queuing delays in the two networks equal to 10 milliseconds. Link bandwidths in wired and wireless networks were set at 100 Mb and 1.2 Mb respectively. We fixed proxy buffer size at 100 packets and varied receiver buffer size in different proportions. As shown in graphs, the latency of file transfer reduced to almost half in case of two split-TCP sessions as compared to a single split-TCP connection.

B. Latency of File Transfer when Queuing Delay in Wireless Link is varied Queuing delay in wireless network is expected to be

more as compared to wired network. In order to observe the effect of increased queuing delays in wireless network, we performed simulations with the following setup. We set bandwidth in wired link as 100 Mbytes and in wireless link as 1.2 Mbytes. The buffer sizes at the proxy and receiver were set equal to 640 Kbytes and 64 Kbytes respectively. The queuing delays in wireless network were increased in different ratios and observations were taken for different file sizes 10 Kbytes to 30 Kbytes. The results are shown in Figure 8 and Figure 9 which show that the latency of file transfer is reduced with parallel paths. As we increase queuing delay in wireless network, latency of file transfer increases in both single and multiple split-TCP. The performance of multiple split-TCP is better than a margin of more than 50% which is an encouraging observation. Noticeable improvement in Staggered-TCP operation is observed when queuing delay in wireless network is more. Figure 6. Graph showing Latency of file transfer vs. Receiver buffer size

for file size 10 Kbytes.

Figure 7. Graphs showing Latency of file transfer vs. Receiver buffer size

for file sizes 20 Kbytes and 30 Kbytes.

C. Latency of File Transfer when wired and wireless Link Bandwidths are varied We varied bandwidths in wired and wireless links in

different proportions and observed latency of file transfer for a single proxy and for two proxies. We performed simulations for file sizes 10 Kbytes to 30 Kbytes.

Figure 8. Graphs showing Latency of file transfer vs. Queuing delay in wireless network for file sizes 10 Kbytes and 20 Kbytes.

Figure 9. Graph showing Latency of file transfer vs. Queuing delay in

wireless networks for file size 30 Kbytes.

The results show significant reduction in latency of file transfer in all cases as shown in Figure 10 and Figure 11. Although latency of file transfer was reduced when Staggered-TCP was implemented, an interesting and positive observation is enhanced performance gain when the difference between link bandwidths in the two domains are more distinct and file size is larger.

V. CONCLUSION We propose multiple-proxies network model for

interconnectivity of wired-cum-wireless networks such that parallel split-TCP sessions are established. The data is striped across proxies connecting the heterogeneous networks. We present a mechanism for TCP management in heterogeneous network across a number of incumbent proxies connecting wired and wireless hosts. Through simulations in ns2 we prove that multiple split-TCP sessions managed in parallel reduces latency for bulk data transfer.

We plan to extend our work to a detailed and robust framework with elaborated role of proxies. We would also

Figure 10. Graph showing Latency of file transfer vs. Link bandwidths (wired and wireless) for file sizes 10 Kbytes and 20 Kbytes.

Figure 11. Graph showing Latency of file transfer vs. Link bandwidths (wired and wireless) for file size 30 Kbytes.

extend our simulations for various file sizes and a larger number of split-TCP sessions across more proxies.

REFERENCES [1] H. Balakrishnan, V. N. Padmanabhan, S. Seshan, and R. H. Katz, “A

Comparison of Mechanisms for Improving TCP Performance over Wireless Links,” IEEE/ACM Trans. On Networking, vol. 5, no. 6, pp. 756-769, 1997.

[2] A. V. Bakre and B. R. Badrinath, “I-TCP: Indirect TCP for Mobile Hosts,” Proc. IEEE ICDCS, pp. 136-143, 1995.

[3] A. V. Bakre and B. R. Badrinath, “Implementation and Performance Evaluation of Indirect TCP,”IEEE Transactions on Computers, vol. 46 no. 3, pp. 260-278, 1997.

[4] S. Kopparty, S. V. Krishnamurthy, M. Faloutsos, S. K. Tripathi, "Split TCP for Mobile Ad Hoc Networks", Symposium on Ad-Hoc Wireless Networks (SAWN in GLOBECOM), Taipei, Taiwan, November 17-21, 2002

[5] F. Xie, N. Jiang, K. A. Hua and Y. H. Ho, “Semi-Split TCP: Maintaining End-to-End Semantics for Split TCP”, in Proceedings of IEEE International Conference on Local Computer Networks (LCN), pp. 303-314, Dublin, Ireland, October 15-18, 2007.

[6] M. Liu and N. Ehsan, “Modeling TCP Performance with Proxies”, Journal of Computer Communications special issue on Protocol Engineering for Wired and Wireless Networks, Vol 27:961–975, June 2004.

[7] A. Sundararaj, D. Duchamp, "Analytical Characterization of the Throughput of a Split TCP Connection", Technical Report 2003-04, Department of Computer Science, Stevens Institute of Technology, 2003.

[8] A. Dunkels, T. Voigt, J. Alonso, “Making TCP/IP Viable for Wireless Sensor Networks”. Work in Progress Session at 1st European Workshop on Wireless Sensor Networks (EWSN 2004), Berlin, January 2004.

[9] A. H. Akbar, K. Kim, W. Jung, A. K. Bashir, S. Yoo, “GARPAN: Gateway-Assisted Inter-PAN Routing for 6LowPANs”, Computational Science and Its Applications - ICCSA 2006, International Conference, Glasgow, UK, May 8-11, 2006.

[10] A. Habib, N. Christin, “Taking advantage of multihoming with session layer striping”, Proceedings of the 9th IEEE Global Internet Symposium (Global Internet 2006).

[11] L H.-Y. Hsieh and R. Sivakumar, “A transport layer approach for achieving aggregate bandwidths on multi-homed mobile hosts”, Proceedings of ACM MOBICOM, Atlanta, GA USA, Sept. 2002.

[12] M. Zhang, J. Lai, A. Krishnamurthy, L. Peterson, and R. Wang, “A Transport Layer Approach for Improving End-to-End Performance and Robustness Using Redundant Paths,” in USENIX, June 2004.

[13] M. Junaid and M. Saleem, ‘Issues of Multihoming Implementation Using FAST TCP: A Simulation Based Analysis’, IJCSNS International Journal of Computer Science and Network Security, VOL.8 No.9, September 2008.