Impact of host mobility in designing congestion control

TCP has a feedback response that can increase its congestion window for successful transmissions, or decrease it for losses. Although, the IP architecture lays down the datagram to tell each packet about their independent routes, congestion control algorithms assume that once a connection is established, all segments follow the same path. To address the issues of congestion control, let’s look at the “network-pipe” model, described by Jacobson. In this model, y-axis distance (pipe widths) represent link bandwidth, and x-axis distance (pipe lengths) represents time (queueing and propagation delays). The data stream flows through one set of pipes to the receiver, and acknowledgements flow through another set of pipes back to the sender.

Based on the network-pipe model, Jacobson’s principles for congestion control can be understood as attempting to keep the network’s pipes full, without overfilling them. To reach equilibrium, the sender should quickly probe the network for a capacity estimate. When a mobile node moves during the course of a TCP connection, the network path between the hosts changes and could give birth to the following issues-

Problem 1: Soft handoffs

For some period of time, acknowledgements for data segments called stale ack sent through the top pipe will be received. Stock TCP uses these stale acknowledgements to clock out data segments that travel through the bottom pipe, and to increase its sending rate. However, this behavior is not in line with the packet conservation principle, as stale acknowledgements do not indicate that segments have left the bottom pipe, where new segments where sent.

Problem 2: Hard handoff

It is possible that an entire window of segments or acknowledgements to be lost. If this degree of loss occurs, then TCP senders may be forced to wait idly for a full retransmission timeout before beginning the process of repairing the losses and recalibrating the congestion window to the new network.

Problem 3: Invalid slow start threshold

The congestion control state of a TCP connection includes a variable, ssthresh (for the slow start threshold), which sets the boundary congestion window between TCP’s exponential and linear increases in sending rate. If, at some point, one host changes connection points, such that the new path, keeping the stale ssthresh established on the old link prevents the new capacity from being used efficiently.

Problem 4 : Network Design, provisioning and stability

If we assume a case where mobility events are considered infrequent over the lifetime of an individual connection, the entire network can still have many within mobile hosts. While a single host’s congestion control behavior will certainly influence the performance of its own connections and others sharing the same links, the design and stability of the network as a whole may be an issue as the number of mobile nodes or connections increases. Figure 2 shows two connections, C1 and C2 that are terminated at mobile nodes R1 and R2. At time t, C1’s segments flow through routers x4 and x1, but at time t+δ, R1 moves from link l1 to link l0. The new path which segments take is through routers x4, x5, and x0. With soft-handoff support, x1 tunnels segments for R1 to x0. The link between these two routers must be adequately provisioned to accept the sudden burst of tunneled traffic, which leads the design of the network to include expensive high-speed links between access routers, which would otherwise not be needed.

Solution:

The Lightweight Mobility Detection and Response (LMDR) algorithm has been proposed as a way of avoiding the TCP problems. Detection of local mobility can be accomplished depending on the underlying network layer and mobility management protocols. A single host is capable of unilaterally detecting its own mobility and locally responding to it by fixing its own TCP state for the half-connection. However, if the half-connection in the reverse direction carries a large amount of data, the remote peer needs to be notified so that it can reset its TCP state.

A new TCP option as shown in Fig3 can be used for this purpose. The first byte identifies the type and the second byte gives its length. The next two bits of the third byte are reserved for future use, and should be set to zero by senders and ignored by receivers. The remaining 6 bits are divided into two 3-bit counters: CNTR and ECNT. A host wishing to use LMDR places an LMDR option on its SYNs. If the remote host supports LMDR, it responds to received LMDR options on SYNs, by placing an LMDR option on SYN-ACKs. Upon connection startup, CNTR is initialized to some random value, and ECNT is left uninitialized until the first CNTR value is received from the remote host. The ECNT field and variable are used to echo received CNTR values. Each time a host moves, it decrements CNTR, and advertises this by continuously transmitting LMDR options on its outgoing segments until it receives back an LMDR option with an ECNT value that matches the local CNTR value. Upon receiving LMDR options, a host sets its ECNT variable to the received CNTR value. The only times when hosts need to send LMDR options are when they are informing peers of their own mobility, or confirming the reception of a peer’s mobility notification. During normal exchange of data between mobility events, there is no need to transmit LMDR options on segments.

Limitations:

The effectiveness of LMDR is limited to situations where end hosts are the only mobile nodes, and the network infrastructure remains fixed. This proposed method might not be very effective if network path changes are caused through attachment to a mobile router rather than the mobility of an end host. However, the proposed LMDR techniques could still be used, if there were some means for mobile routers to notify attached end hosts of mobility events.

The functioning and implementation along with formulation have been taken from the references listed below.

REFERENCES

• Host-to-Host Congestion Control for TCP, Alexander Afanasyev, Neil Tilley, Peter Reiher, and Leonard Kleinrock, UCLA, IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 12, NO. 3, THIRD QUARTER 2010

• JACOBSON, V. Congestion avoidance and control. In Proceedings of SIGCOMM ’88 (Stanford, CA, Aug. 1988), ACM.

• A Transport Layer Approach to Host Mobility, MOBICOM’02, September 23-28, 2002, Atlanta, Georgia, USA. Copyright 2002 ACM 1-58113-486-X/02/000