| Summary: | 碩士 === 國立清華大學 === 資訊系統與應用研究所 === 96 === In the task of melody recognition, CBMR (Contend-based Music Retrieval) system searches for a query song by extracting features from the acoustic input from the user and comparing them with the songs in the MIDI database. Due to the computational complexity of similarity comparison of song features, the performance is limited under the single server settings. In order to alleviate this problem, parallel computing was invented to increase the computation efficiency.
If a large number of users are accessing the parallel computing system simultaneously, the overall response time would increase with the increasing number of users. This thesis discusses several solutions to the above problem, including a Round Robin strategy to distribute the user requests to several slave servers, mode switching consideration between single request and simultaneous multi-request, and comparison computation reduction techniques. In general, by using the parallel system with the queue strategy (5 slave servers) and adopting linear scaling and searching-from-beginning techniques, the average response time for 100 simultaneous requests is 24 seconds. Under the new framework of Round Robin system, the average response time can be reduced to around 10 seconds.
On the other hand, this thesis also discusses an approach to accelerate the recognition process based on clustered PCs. It updates the threshold distance between the query song and the compared song in real time by exchanging information between master and slave servers. The current calculation is aborted once the partial computation result is larger than the threshold distance and the comparison for the next song is started immediately. Thus, the time is saved for unnecessary computation and the efficiency is increased.
|
|---|
