| Abstract: | Today's web servers must have the ability to deal with large data sets, and their performance mainly depends on the control mechanism of the disc cache. Though the cache replacement algorithms in operating systems generally perform well, application-specific policies can often perform much better. In this paper, we present the application-controlled caching in user space (ACCUS) mechanism for web server disc caching. With this mechanism, an application can schedule the service requests based on the cache status of the requested files to improve its performance so that the system can gain a high parallelism of CPU processing, networking I/O and disc I/O. An application can service the cached files with a higher priority to the ones not cached, which helps reduce the latency caused by disc I/O blocking. Meanwhile, the application can enforce a policy for domain-specific cache management to obtain a higher cache hit ratio. ACCUS is implemented in two web servers of different models, Flash and userver. Despite the great differences in architectures, the empirical results show that both servers can achieve high throughput under heavy workloads with ACCUS enabled. Performance analysis for two other typical web servers, Apache? and Zeus Web Server are also conducted and the results indicate that ACCUS can achieve significant improvement in performance. |
