TOWARDS EFFECTIVE MULTI-TASK ADAPTIVE-TICKS TIMER IN LINUX KERNEL FOR FULL TICKLESS FUNCTION

Abstract

Scheduler timer architecture has a significant impact on operating system performance and power consumption. The current generation of Linux kernel supports multiple timer implementations, including periodic ticks, Dyntick-idle, and Adaptive-ticks. The Adaptive-ticks kernel offers the benefits of previous generations, with additional improvements to performance and power consumption. Adaptive-ticks is a relatively new feature included in the Linux kernel since version 3.10. In the current Adaptive-ticks implementation, the scheduler will check whether there is only one task in the ready queue and then reprogram the HR timer to run without interrupt for 1 second before hitting the next timer tick. Adaptive-ticks kernel eliminates the need for running timer interrupts at regular intervals, thereby letting the CPU execute the workload without interruption. This also means fewer cache line replacements for executing interrupt and fewer CPU cycles spent on executing timer interrupt service routine. This feature has not previously been studied in detail and it was one of the preliminary focus of this research to benchmark the performance improvement achieved by this feature so as to understand the impact on application performance and system behavior. Since power consumption is a major motivation for the implementation of Adaptive-ticks kernel, the current state of research on the feature will remain incomplete without benchmarking the power consumption of the feature. As a second step, we studied the power consumption of Adaptive-ticks kernel by benchmarking it in typical CPU, IO, and memory-bound workloads by using In-bound power measurement tool, synthetic benchmarks, and real-world workloads, to demonstrate the impact on power consumption. In the Adaptive-ticks kernel, there is no support for multiple tasks, meaning support is limited in scope to situations where only one task is available in the ready queue. Since most modern applications implement parallelization to no small extent, this limits the current effectiveness of the Adaptive-ticks feature. Application optimization is one technique that could help overcome these limitations. Therefore, we proposed and implemented an algorithm to improve application performance and reduce power consumption by reorganizing worker threads in an application and benchmarking the solution to quantify the enhancement in performance and power consumption. The current Adaptive-tick feature is limited by design to support only one task in the ready queue. To solve this limitation in the design, we designed and implemented the Full Adaptive-ticks feature, which supports up to eight tasks in the ready queue. This implementation was a necessary step forward in terms of improving application performance and system power efficiency. The implementation includes an algorithm to dynamically calculate the time slice required by each task in the ready queue. Rather than a set quota of 1/CONFIG Hz seconds, the approach would be to have the time slice be equal to 1/(number of tasks ready). In this method, for example, when there is only one task in the ready queue, just as in the current Adaptive-ticks kernel, the task gets one second of the time slice and there is only one timer interrupt per second. However, in the modified design, if there are two ready tasks, the time slice will be 500 ms and there will be two scheduler timer interrupts per second. When the number of tasks increases, we will see as many numbers of interrupts per second as the number of tasks in the ready queue.