The Effects of Mobile Agent Performance on Mp3 Streaming Applications

Extended Abstract

There are many areas in which mobile agent technology can be deployed. Researchers are now looking at this technology as a solution in their respective research areas, such as information retrieval, distributed computing, and network management. This increase of interest is due the basic characteristics of mobileagents - autonomous and mobile. In the paper, we are exploiting the abovementioned characteristics of mobile agents, and applying this technology in Internetenvironments. Using a mobile agent as a transparent content transcoding proxy in a proxy based network architecture, we show how a mobile proxy agent can be used to improve quality of service provided to the end user connected via heterogeneous access networks, using computing devices which have very different capabilities. In our framework, we are using software agents, both static and mobile, programmed with different proxy functions, to provide transparent content transcoding and network enhancements to the end user. This is an extension to the current proxybased network architectures [1], which uses dedicated and centralised proxy servers to provide the transparent transcoding. Using mobile agents make our framework more efficient, flexible and scalable than the static proxy architectures, as a proxy agent can follow a roaming user, execute on a lightly loaded server, and/or optimally located to keep the number of hops between itself and the end user to a minimum. There have been different performance measurements on various mobile agent platforms available today. Narasimhan [3],[4],[5] performed two experiments on IBM aglets to determine the effect of network congestion and the size of the aglet affects migration time. The first experiment was measuring the round-trip time of an aglet with increasing size, and the second experiment was measuring the round-trip time of an aglet when it travels around the world via various hosts. The results provide us with a rough estimate of the performance of the IBM aglets; however, due to the simplicity of the experiments, we cannot draw any conclusions on the suitability of aglets for our framework from this work. Silva et al. [7] extensively benchmarked 8 different mobile agent platforms: IBM Aglets, Concordia, Voyager, Odyssey, Jumping Beans, Grasshopper, Swarm and James. They performed 12 experiments, testing each mobile agent platform on execution time, migration time, the amount of network traffic and robust- ness. They concluded that James provides the best performance out of the 8 mobile agent platforms they evaluated. We can use this study as a guideline on which mobile agent platform has the best performance, but we cannot use this study to decide the viability of the use of mobile agents in our framework, because the experiments Silva et al. performed used unrealistic scenarios. Moreover, the study of Narasimhan and Silva et al., were aimed at measuring the performance of mobile agents, as a result they were not consider the specific application of the technology. Therefore the results they presented are only the performance of the mobile agent in a general case. To evaluate the viability of our mobile agent based proxy architecture, we developed a prototype mp3 audio streaming application using IBM aglets. Our experiment concentrated on the migration time of the aglets, and the perceived quality of the MP3 stream without any transcoding, and with a mobile aglet transcoder. The prototype used a modified version of Obsequieum [6] as a streaming mp3 server, and Freeamp [2] as a streaming mp3 client. Unlike other streaming mp3 applications, which use TCP as their streaming protocol, Obsequieum streams mp3 audio using RTP. To test the performance of the system we conducted two experiments. The aim of the first experiment was to determine the amount of audio lost when the proxy agent migrates from one host to another. The results we obtained from the experiment agreed with our previous streaming audio emulation results, that the number of audio packet loss is directly related to the network traffic. However, since RTP contains time stamping, we discovered that on average, the minimum amount of audio loss is 240ms. Under heavy network traffic, the amount of audio loss can be as much as 1700ms. We also discovered the effect of the client application with buffering and without buffering. When buffering is on (we used the default value of 3 seconds), at minimum network traffic, the user only hears a slight skip of audio - the 240ms loss of audio. However, as the buffer empties due to the high frequency of migration, the application stops playing the audio for 3 seconds for re-buffering. From the end user’s perspective, s/he experiences silence for 3 seconds and still suffer a 240ms loss of audio. The situation is more severe under heavy network traffic. As the application buffer appears to be emptied at every agent migration, the application has to re-buffer. This causes a period of silence for 3 seconds and a loss of 1700ms of audio. The second experiment was aimed at investigating how a proxy agent can minimise the loss of audio. A test network is set up, which consists of 3 different subnets. The mobile client can roam between subnet B and C, and the proxy agent can follow the mobile client by residing on base station B and C. Subnet B is a congested network. The experiment involved the mobile client roaming between subnet B and C while the server is streaming mp3 audio via a mobile proxy agent. We are hypothesising that if subnet B congested, the amount of audio loss will be high. If the mobile client roams to subnet C, but the mobile proxy agent remains in base station B, the amount of audio loss will be higher, since the RTP packets has to go through the congested network twice. However, if the mobile proxy agent migrates to base station C, there will be a period of which there will not no audio, but once the migration process completes, the amount of audio loss should be less than before, as the audio packets are no longer route through subnet B. The initial results show that the above assertion is valid, and that the overall perceived quality is improved.