Project description: Multimedia data, such as image and video, has become one of the most popular data types being processed every day. Due to the prevalence of multimedia retrieval applications and their prohibitive execution times, it is necessary to understand their performance characteristics for evaluation and optimization. To achieve this goal, we build up MMRBench, a public available benchmark suite containing representative state-of-art multimedia retrieval applications, including original version, our implemented POSIX version (thread-level parallelized) and map-reduce version (task-level parallelized). We also do study on these applications about their architectural characteristics as well as other performance related analysis such as input sensitivity, memory/computation intensity, floating operation sensitivity and potential thread-level parallelism. Based on the suite of MMRBench and automatic tools we provided, it is easy to construct a real multimedia retrieval system or do research on related architecture, such as system evaluation, architecture design and accelerator.