9 Conclusions and Directions for Future Work

 

The main goal of this MQP was to develop an improved method for transporting real-time, multimedia data over the internet. Timed obsolescence provided a means for achieving our goal, but necessitated an entirely new transport layer protocol (XUDP) for a complete implementation. This task was completed, producing an extension to the User Datagram Protocol, with flow control mechanisms similar to that found in the Transmission Control Protocol. As evidenced by the results presented earlier, XUDP successfully preserves the linear time relationships of real-time multimedia data under a variety of network conditions.

However, work remains to be done regarding the optimization of XUDP and the preparation for eventual distribution as a unique contribution to the growing array of multimedia transport layer protocols. As of this stage, much of the code comprising XUDP is designed for easy modification, readability and expansion and is not optimized for speed. Below are some items that can be cleared up in future project work:

• Minimize context switches in the main loop. XUDP is constructed around a series of queues that are searched for data to process by a number of functions accessed sequentially from a main program loop. Quite often, the functions are called only to determine that there is no work to be done, and control is returned to the loop. If the queues are checked for data to be processed on the main loop level, costly function calls can be avoided and stack times will be significantly reduced.
• Making use of hash tables to interface with queues. Currently in all cases, XUDP uses a linear search through linked-list nodes to locate information. It may be worthwhile to investigate the possibility of the creation of dynamic hash tables that would significantly reduce the amount of searching necessary.
• Less costly RTT statistical calculations. The round trip time statistics are calculated with floating point arithmetic on a per-packet basis. Congestion Avoidance and Control[19] outlines a method of using integer arithmetic on scaled powers of two, which is cheap and effective on any platform.
• Theoretically derived RTO and window sizing algorithms XUDP is using the retransmission timeout with variance estimation algorithm first developed for 4.3BSD Tahoe TCP. Worse, the congestion anticipation window sizing algorithm currently responds too quickly to changes in round trip times. Some time can be spent determining more suitable solutions to these problems.
• IPC with shared memory. Parcels are being sent from the user's application to the local XUDP server through UNIX streams. If shared memory handling functions were added, some protocol stack delay time could be eliminated.
• Implementation directly on top of IP. XUDP is dependent on UDP for its transport mechanism. An implementation with raw sockets built on top of IP would place XUDP at the same layer as TCP and UDP, further reducing stack times.