Controlling Kernel Scheduling from User Space: An Approach to Enhancing Applications' Reactivity to I/O Events

In this paper, we present a sophisticated mechanism that allows an application to tightly control the way I/O events are handled within the underlying operating system’s kernel. The goal is to provide an efficient user-level interface to allow applications to quickly detect and handle asynchronous I/O events. Reactivity is obviously a crucial property for many time-critical applications. For instance, it is a major issue in distributed applications where poor reactivity to communication events may have a direct influence on the observed latency of communications, and may even impact on the whole application behavior. Our approach is based on an extension of the Scheduler Activation kernel scheduling model, originally proposed by Anderson et al. The idea is to minimize the time elapsed between the detection of an I/O event within the kernel and the scheduling of the appropriate user-level thread within the application. The progress of I/O operations still takes place in kernel space, but the scheduling decisions are taken in user space. We have implemented the core of our proposal within the Linux kernel, and the PM2 multithreading environment was ported on top of the corresponding user-level API. Our experiments demonstrate that the implementation of our kernel extensions is very efficient and that the overhead introduced by the user-level interface is very small. Moreover, the observed results show that our approach competes favorably against approaches based on the use of signals or kernel threads.