Problems on TCP

The well-known advantage of TCP protocol is that TCP can guarantee the integrity of the transmitted files. Anything might take place on every network path. TCP concerns about the status of the connection. Delivering packets on paths with abnormality makes the performance to degrade.

The event that TCP most concerned is packet loss. When packets got lost, TCP assumes that there might be congestion in the network. It then tries to prevent the network traffic from becoming worse so it slows down the data sending rate with various kinds of mechanisms. In general, packet loss affects TCP performance a lot.

TCP detects the network abnormality status on the connection in advance. The long propagation delay might mislead TCP protocols to assume that the packet was lost. The propagation delay is a part of RTT which is also a reference of timeout. The long propagation delay affects the value of timeout.

On the Internet, there are may be multiple routing paths for a connection and the routing information is always changing. A file may be divided into pieces and carried by TCP packets.

(5) (6) (7)

17

The packets might be traveling with different path depending on the changing routing information that is kept in intermediate-routers. Packets traversing different paths may arrive at the destination out of orders. The order of incoming packets on the receiver and the order of incoming acknowledgements might be out of order. TCP takes the packet with missing serial number as being lost and then starts to retransmit the lost packets to the destination. This kind of retransmission will hold the size of cwnd or decrease the cwnd and consume the bandwidth.

2.3.1 Packet Loss 

All packets injected to the router race the limited bandwidth to be transmitted to the link.

No matter what kind of queuing policy adopted, there are always packets dropped in the router node. The reasons might be that the node is not powerful enough to process numerous packets, the incoming or outgoing link might have bad signaling or interference, or the outgoing bandwidth is not sufficient. Also, this might be caused by link failure.

When TCP protocol encounters the packet loss, it has to decrease the cwnd size to slow down the traffic injecting to the network. Every TCP protocol has its own algorithm to determine what to do. Fast retransmission and fast recovery will be applied to retrieve the few lost packets. Fast retransmission determines how much to eliminate and fast recovery determines how to retrieve the state as it is. Anyway, the data rate of TCP decreases. Even TCP SACK can recover multiple packets loss but the cwnd is temporarily frozen. The recovery period costs some packets to be re-sent. From total throughput concern, it is still lower than expected.

If the packet is timeout, TCP protocol has to enter slow start phase to reduce cwnd size to as small as possible. To release the loading from the path, TCP protocols did its best to re-start the transmission from one packet and allow the path become healthy later. For heterogeneous environment, the paths are coexisting independently.

18

2.3.2 Long Queuing delay 

The TCP throughput is affected by RTT [1] a lot. If the end-to-end path to deliver the TCP packets features long queuing delay, TCP performance will be poor. Theoretically, the intermediate-nodes will have the best routing information for specific destination. But this kind of routing is not fair. All traffic will go through the selected path and there may be no traffic on the other paths which has smaller bandwidth. If the traffic is too concentrated, the queuing delay of the node would be long. Unfortunately, current routing policy does not implement the load balance mechanism among the connected paths. Traffic crowded in certain specific routing path is very common. The RTT of TCP packets is generally long in the real Internet and the performance is degraded due to the long RTT.

2.3.3 Different Routes   

On the Internet, the intermediate-routers are linking to each other to construct a physical network. There are numerous paths linking a source and a destination. There are also various kinds of static routing configuration and dynamic routing protocol, such as OSPF, RIP, running among the routers to exchange the routing information. That means the route from source to destination for this packet might be different with other packets. Some nodes may be visit commonly by most packets in a connection, but some nodes may be only visited by few packets in the same connection. Among different paths, propagation delays are also different. The status of the paths is also concerned. A path might be congested and the packets going through the path is likely to be dropped. Another path might be selected temporarily due to its narrow bandwidth that enlarges the round trip time. There might also be a path that has poor linking signals due to miscellaneous reasons, such as a path whose outgoing path is disconnected or down.

19

All situations lead the TCP packets to a result – sequence numbers are not consecutive.

No matter the sequence order received by receiver or the order of acknowledgment packets received by the sender, the transmission goes abnormal.

As we mentioned above, sequence numbers being not consecutive will make the sender to guess what happened to the path. And the TCP sender would take the corresponding actions to face the problems. The TCP cannot select another path to go even the TCP is looping in the recovery phases. There are multiple routes ready to transport, but this is not helpful for TCP to recover the congestion condition.