ECEM: an event correlation based event manager for an I/O-intensive application

Abstract

In Internet servers that run on general purpose operating systems, network subsystems and disk subsystems cooperate with each other for user requests. Many studies have focused on optimizing the data movement across the subsystems to reduce data copying overhead among kernel buffers, a network send buffer and a disk buffer. When data are moved across the subsystems, events such as read requests and write requests for data movement are also delivered across the subsystems by the servers and the operating system. However, there have been fewer studies on the optimization of event delivery across the subsystems. In conventional operating systems, an event from a disk subsystem is delivered to a network subsystem regardless of the status of the network subsystem. If the network subsystem is not ready for data sending, the execution of the server will be blocked, which causes scheduling and context switching overheads. This non-contiguous execution will incur deficiencies such as avoidable process blocking, context switching, cache pollution and long response time. To alleviate the deficiencies, we have developed inter-subsystem event delivery mechanisms that define event dependencies among the subsystems involved. We define an event correlation based on the happened-before relation. We propose deferred event delivery (DED) and disk-to-network splicing (DNS) to suppress scheduling and context switching during I/O request processing. We performed experiments on Linux 2.4 and the experimental results show that the number of context switching is reduced by up to 20% and server data transmit rate is improved by 4.0–8.1%.