- Split TCP recommendations are not suitable for networks containing asymmetric routing. To apply the split TCP solution, you need to ensure that the mobile device is unblocked in the intermediate media node, and for some networks, this either cannot be done, or the intermediate media node is retracting from the wireless network boundary. Complete, this is actually unable to practice, and may cause routing inexplica (non-Optimal Routing).
- Split TCP, as indicated by its name, no problem with UDP is unable to resolve any issues related to UDP.
Note: Use split TCP does not need to reject concurrent use of the IP end-to-end connection. The correct split TCP usage is to manage and control each application or each connection, so that the user can decide whether to use a specified TCP connection or application to use split TCP, or on IP End peer operation.
Recommendation: Split TCP recommends that the TCP syntax is changed, and it is not recommended. In addition, some custom protocols on wireless connections such as MOWGLI, etc., are not recommended, because their claims: (1) Improved TCP rather than redesign; (2) Alternate end-to-end sessions are allowed at any time.
4.10.2 Application Level Proxies
Today, the application level agent is widely used in the Internet, including: Web Agent Cache, Relay MTAS (Mail Transfer Agents, Mail Transfer Agents), and secure transport agents (such as SOCKS). In the "Split TCP connection", the application level agent acts as an intermediate medium. Thus, some problems in the wireless connection, such as a combination of crowded wide-area Internet channels and wireless LTN connections, automatically extended to other areas.
Application protocols often use a large number of round-trip cycles, a large number of headers and low efficiency encodings, even in a routine application protocol operation, there may be many unnecessary data may also be transmitted to wireless connections. In most cases, this overload can be easily alleviated by running an application level agent at the intermediate medium node. Under the LTN connection, the enhanced manner specified by the application can inform the application level agent, which can significantly improve performance. This agent may use an enhanced version of the application protocol on a wireless connection.
Montenegro, et al. Informational [Page 26]
In the LTN environment, the enhancement of the application layer will provide more significant performance than any transport level enhancement.
The MOWGLI system provides full support for Agent-Proxy PAIRS, composed of Agent-Proxy PAIRS, and the proxy server at the mobile device and the proxy server at the intermediate media node. This pair can be known for the application, but it can be fully transparent, but in all connection time, it is under the control of the end user. See MOWGLI WWW [Laklr95], [LHKR96], and WebExpress [HL96], [HL96], [CTCSM97] for application specified agents.
Recommendation: Use the application level agent if there is conditional, that is, the application must support the agent; to apply the selected agent to the user's control.
4.10.3 Discipline and its derived (Snoop and ITS DeriVATIVes)
Berkeley's Snoop Protocol [Snoop] is a product of the connection layer reliability mechanism and the split connection scheme mixed. It is better than the split TCP scheme because it retains the end-to-end syntax. Snoop is doing the following:
1. Topically (wireless connection range) relay the lost packet, rather than allowing TCP to be implemented by end-to-end mode.
2. Inhibition is re-confirmed by the receiving end to the sender, thereby avoiding congestion avoidance avoidance avoidance avoidance avoidance avoidance.
Thus, the design purpose of the SNOOP protocol is to avoid unnecessary fast relays of the TCP sender when the wireless connection layer is partially relayed. It is assumed that a system does not use the Snoop agent: TCP sender S sends an information package to the receiver R via an intermediate media node in. It is assumed that the transmit party sequentially transmits the packets A, B, C, D, E, and pushes it to the wireless receiving end R by IN. If the bag B is lost, the intermediate medium node is then relayed by the intermediate medium node, in which case the package received is second, C, D, E, and B. Repeated confirmation signals will be triggered after receiving packages C, D, and E. When the TCP sender receives three duplicate confirmation signals, it triggers fast relay (which will raise the relay, and the reduction in the congestion window). Although the link level relay is underway, the fast success can still occur, thereby reducing the throughput efficiency.
Montenegro, et al. Informational [Page 27]
Snoop [snoop] Discard the corresponding TCP repetition (Dupacks) at the intermediate media node to resolve this issue. At the intermediate media node, try the approximate listening of delayed repetition (DELAYED DUPACKS, see Section 4.5). This solution is only useful when the possibility of the fast relay caused by wireless errors is not negligible. Special circumstances, if the wireless connection uses the protocol of the stop-and-GO mode (or sequentially transmitting the package), this solution is not very effective. Similarly, if the bandwidth delay product of wireless connection is less than four segments, the intermediate medium node sends three new packages in the packet loss to be relayed. Because the trigger is triggered, at least three repeating packages are required. When the wireless bandwidth delay product is less than four packages, the scheme such as Snoop and delay is not useful (unless there is a problem with the link layer design) . Conversely, when the wireless bandwidth delay product is sufficiently large, SNOOP can exhibit significant performance improvement (compared to standard TCP). For a deeper discussion on this topic, see [VAIDYA99].
The delay recurring package can play a performance improvement in a good environment in Snoop. Typically, performance can be improved by delayed repeating package schemes because it has the function of reducing the packet loss rate caused by congestion and relay errors. When generated with congestion-related packet loss, the delay relay will perform unnecessary delay relay, and thus the delay relay will not improve performance like SNOOP when processing congestion packet loss. However, the simulation results [VAIDYA99] show that the delay relay can achieve significant performance improvement in the case of moderate congestion packet loss.
Similar to Snoop, WTCP [WTCP] also retains the side-to-end syntax. In the WTCP, the intermediate media node uses a complex scheme to hide it to restore the wireless connection part of the error (including the relay caused by the error recovery, or the time for processing congestion). This idea contributes to the sender to better estimate the round-trip time period. In order to work more efficient, it assumes that the TCP endpoint (endpoint) implements the timestamp option in RFC1323 [TCPHP]. However, unfortunately, the TCP supported by RFC1323 is currently not common. Moreover, the WTCP requires only modified at the intermediate media node. SNOOP and WTCP require intermediate media nodes to check and operate communication between portable mobile devices and TCP servers on the wired Internet. Unless the intermediate media node also shares security connections between mobile devices and their end-to-end connection points, Snoop and WTCP cannot work when IP communications is in an encrypted state.
Montenegro, et al. Informational [Page 28]
They also require data and corresponding confirmation information to cross the same intermediate node. Moreover, if the intermediate node performs relaying the information packet through the wireless connection in the transport layer, the link layer will cause a double impact on the link layer. Snoop designers have made this, that is, there is a link layer of TCP-aware. This is the correct solution: the connection layer and the network layer should enhance communication with each other, not the dead plates specified in the traditional OSI model.
Regardless of the transmission mode or channel mode, the IPSec's ESP header encrypted IP enrollment causes the intermediate node to deal with the TCP header and valid. This eliminates the possibility of Snoop (and WTCP) applications because it needs to check the TCP header in both directions. Possible solutions include:
- Establish security relationships with customers and servers for the intermediate nodes of Snoop (or WTCP).
- IPsec channel mode is terminated at the Snoop intermediate node.
However, these techniques require user trust in the intermediate node. Users must protect their privacy and want to improve performance using SSL or end-to-end security mode SOCKS, which is already implemented in the transport layer. But they are not like IPsec to resist certain security attacks (eg, guess attacks based on TCP sequence numbers).
Recommendation: Realize Snoop immediately at the intermediate node. The results of the study are delightful. It is invisible to optimize, and no changes are required for clients and server side, but as long as the intermediate node is modified (for basic SACK Snoop). Of course, as mentioned above, Snoop is still less efficient in the following cases:
1. Wireless connection provides reliable, sequential package delivery;
2. The bandwidth delay product of wireless connection is smaller than four sections.
4.10.4 Use Performance Enhancement Agent Processing Disposable Phase (Peps to Handle Periods of Disconnection)
Connection interrupt sites often occur in wireless networks, whether it is a handoff state, or due to resource lacking, or some normal obstacles may cause connection interrupts. In the interrupt state, the TCP sender cannot receive the confirmation response it expects. After the succession period is full, TCP will close all congestion windows related to the error. When the connection is interrupted, the relay package is useless.
Montenegro, et al. Informational [Page 29]
[M-TCP] is committed to helping TCP can better handle switching and disconnect state, while also retaining end-to-end syntax. M-TCP adds some elements: Adding Super User Hosts (Supervisor Host, SH-TCP) at the boundary of the wireless network. The intermediate node monitors the communication status of the sender to the mobile device, which does not abort the end-to-end syntax because the confirmation signal (ACKS) sent by the intermediate node to the sender is exactly the mobile node. The principle is not confirmed to the last byte. Therefore, the SH-TCP can turn off the sender window by transmitting a window set to zero as a confirmation of the last byte, so that the sender can maintain a persistent mode.
The second optimization needs to do some work between intermediate nodes and mobile hosts. For the latter, TCP needs to pay attention to the current connection status to freeze all timers during the disconnection. When the connection is restored, the mobile device sends a confirmation signal (ACK) that is marked with the most byte number received. The intermediate node monitors the communication flow of the wireless connection. When it assumes that the mobile device is dropped, a notification will be issued to the SH-TCP, and a confirmation signal is transmitted by the SH-TCP to close the sender window (as described above). After the intermediate node receives a repetitive confirmation of the tagged connection, it knows that the mobile device is coming back, then it will issue a repeated confirmation signal (ACK) to the sender and open the window. The sender will exit the persistent mode, returned to the previous state, and transmitted data with previous rates. The sender will relapse any data confirmed by the previously not confirmed. It is not a relationship as long as it is not overwriting or software, because the previous intermediate system can shorten the window and can modify it with a new window after it receives the instructions from the mobile device.
Recommendation: M-TCP is not currently the current adoption program, because there are still many better recommendations, and current TCP / IP implementation has some problems in processing zero windows. Please continue to pay attention to research in this area.
4.11 Title Compression Selection (Header Compression Alternatives)
Since the remote narrowband network (LTN) is bandwidth constraint, it is worth testing for each byte of the public segment.
For the definition of TCP and IP head compression, see [RFC1144, IPHC, IPHC-RTP, IPHC-PPP], which provides the following benefits:
- Improved the response time of interactive applications.
- Allows the batch data to be delivered with a small package at a good linear efficiency.
Montenegro, et al. Informational [Page 30]
