The combination of ATM and satellite technologies working with TCP/IP, forms an internetwork architecture that has the potencial to provide seamless networking capabilities. The three major service categories for TCP/IP data transport over satellite are ABR, GFR and UBR. A comparison among these service categories based on the implementation complexity, buffering requirements for switches and ATM end systems, network bandwidth utilization and bandwidth allocation will be showed in this paper.
ABR (Available Bit Rate) services guarantee a Minimum Cell Rate (MCR) to the VCs and a fair sharing of any unused capacity. ABR uses a rate-based closed-loop feedback control mechanism for end to end congestion control or, break it into several hops by using the Virtual Source/Virtual Destination option (VS/VD). On a TCP connection over ATM using ABR service, the problems come from the interaction between both congestion control methods, because they use different ways to detect congestion in the network. With the current implementations of TCP, ATM network is limited to packets drops as the only way to advise the TCP sources of the congestion. Another problem with TCP over ABR is the long RTT (Round Trip Time) in the satellite environment. This service requires a resource allocation algorithm for CAC (Connection Admission Control) and UPC (Usage Parameter Control). The VS/VD option increases the complexity of the mechanism.
GFR (Guaranteed Frame Rate) service requires the specification of a MCR and a Maximum Frame Size (MFS) for each VC. For satellite ATM networks, TCP congestion control mechanism has more effect on TCP throughput than on ATM buffer management policies. The GFR is adequate to be used in applications that don’t need traffic parameters for a VBR VC (Virtual Channel) and capability for rate-based feedback control. This service has a similar complexity compared to the ABR service.
The UBR (Unspecified Bit Rate) category service is more commom to be used over satellite ATM networks, because this service doesn`t include flow control and hence it depends on transport layers to provide flow control. When TCP uses UBR, and cells are dropped at the ATM layer, TCP has to recover by the ACK information using its congestion mechanisms. When a cell is dropped by the source switch, the destination drops the entire packet. TCP then timeouts and retransmits the entire packet. A minimum GR (Guaranteed Rate) will be used by the UBR traffic to protect the UBR service category from the total bandwidth starvation, and to provide a continuous minimum bandwidth guarantee. The UBR basic service has a simpler implementation complexity because it does not require any CAC and UPC. For satellite ATM networks, the TCP congestion control mechanisms have more effect on TCP throughput than ATM buffer management policies therefore a simple Tail Drop is enough to be used. The MCR (Minimum Cell Rate) guarantee requires a design option like a scheduling mechanism which select one of the queues at each scheduling time. The aquisition of this service is still very expensive and this extra cost for the per-VC queuing is not a good idea to satellite ATM networks using onboard switches.
The performance of TCP over satellite ATM can be revealed very poor due to long RTT, high bandwidth, buffer size, bandwidth asymetry, channel access netwok interactions and BER (Bandwidth Error Rate) characteristics. In order to enhance the performance of the TCP in the Internet, several algorithms and options have been created.
The goal of this paper is to show the ways to provide a significant improvement of the throughput and delay of the standard TCP/IP applications, identifying elements which limit the performance over a satellite link and proposing possible solutions. The main challenges that must be faced and the suggested solutions are described below.
- TCP window must be able to reach large values in order to efficiently use the available bandwidth. This is possible using the window scale factor in the TCP header that gives a credit allocation in octets. When the window scale option is in use, the value in the window field increases. This solution increases the TCP efficiency if a low loss of segment occurs at each window size transmission. Otherwise, for the UBR and GFR where the system doesn’t know the network congestion and send packets into the ATM network as fast as it is possible, the losses will make TCP return to the Slow Start phase many times. This results in a low efficiency. The ABR service using the feedback control mechanism knows the situation of the ATM network so that large buffers in the edge routers can be used when the network is temporarily congested.
- The buffers size must be well dimensioned to avoid overflow during the Slow Start phase before reaching the network capacity otherwise, this will result in a TCP performance deterioration. Several studies show that a buffer size equal to about half the round trip delay–bandwidth product (BDP) of the TCP connections, provides a high TCP throughput over satellite-UBR. With ABR, satellite gateways can buffer large amounts of data, while the buffer requirements of the onboard ATM switches can be minimized.
- TCP considers the loss of packets as a signal of network congestion and consequently, reduces its window. This results in a severy throughput deterioration when packets are lost for other reasons than congestion. Non-congestion losses are mostly caused by transmission errors. Trying to reduce the loss of packets it is proposed the use of FEC (Forward Error Correction) mechanism to improve the quality of the link. FEC provides a BER less than 10-11 , in the worst condition. The reason to the low rate loss is that some corrupted packets can be corrected directly without needing a retransmission and it does not interfere in the TCP mechanisms. Therefore this technique has been used in a satellite ATM network.
Using FEC, this paper proposes a comparative evaluation of the TCP implementations like Tahoe, Reno, New-Reno and Vegas for the integration of satellite links. Some preliminary results allow us to conclude that a higher efficiency of TCP over satellite ATM network using UBR service category was reached by Vegas. Differently of the others implementations which face packets losses as a signal of network congestion, Vegas uses the difference in the expected and actual rates to adjust its window size. This fundamental difference in the bandwidth estimation schemes enables TCP Vegas to control the network congestion. Vegas uses the most accurate RTT estimate to detect the congestion and it can retransmit a packet even before the receipt of three duplicate ACKs.
