Performance Evaluation of the mmstream Implementation

    Performance tests were carried out on Sun UltraSPARC workstations, using the Java interpreter contained in Sun's Java Development Kit 1.0.2. The test application used Byte_Producer objects to generate media chunks, RTP_CoDec objects for encoding and decoding the chunks, and UDP_Multicast_Connection objects for forwarding and reception packets. The maximum numbers of produced or consumed chunks per second achieved with this configuration are shown in Table for different sizes of the Chunk objects' byte array.

 
Table:  Performance of the mmstream implementation. The Chunk size is given in kB, whereas the respective performance for the Producer and the Consumers is measured in Chunks per second.

The observed chunk rates are lower than required for media such as full-motion video. This behaviour has two reasons:

• Interpretation. It must be kept in mind that the above chunk rates were achieved using interpreted Java bytecode. However, interpreting bytecode is slower by orders of magnitude than executing native machine code. The situation will improve with the advance of  just-in-time compilers, which were not included in the JDK 1.0.2. Just-in-time compiler will be integrated into Java interpreters to generate native machine code from bytecode on the fly. Sun [GM95, Sec. 8.1,] claims, that the performance of machine code generated from Java bytecode will turn out roughly equivalent to the performance of machine code compiled from C++ source code in traditional ways.
• Implementation style. The current implementation puts the main emphasis on gathering and displaying performance and network state data. For example, on reception of each RTP packet the complete GUI component for the corresponding RTP_RemoteSource object is updated with newly computed reception statistics information. This strategy entails a considerable performance penalty.