Time Donating Barrier for efficient task scheduling in competitive multicore systems

Abstract

Nowadays, co-locating multithreaded applications on a multicore system has increasingly become a common case in cloud data centers, where multiple threads generally compete for computing resources. These competitive environments may suffer problems of system throughput and fairness caused by barrier operations in multithreaded applications. This is because most implementations of the barrier synchronization are based on the spin-then-block mechanism in which spinning–waiting threads probably waste computing resources and relinquish cores to other co-running applications after they are blocked. This paper attempts to find a new and intuitive way to improve the efficiency of barrier in competitive environments, and answer the question: Can we leverage the timeslices of waiting threads to accelerate barrier operations?Targeting this question, we propose a novel barrier synchronization mechanism named Tidon (Time Donating Barrier). The basic idea of Tidon is to donate the timeslices of waiting threads to their preempted, laggard siblings in order to accelerate barrier operations, different from traditional static spinning and blocking. We implement Tidon based on the GNU OpenMP runtime library (libgomp) and Linux kernel with new, lightweight system calls. Our evaluation with various sets of co-running applications demonstrates that the advantages of Tidon include (1) alleviating the performance degradation of barrier-intensive applications (e.g. improving the performance by up to a factor of 17.9 and 2.3 compared to the default barrier implementation of OpenMP in Completely Fair Scheduler and Balance Scheduling, respectively) while not hurting or even improving the performance of non-barrier-intensive applications, and (2) maintaining good fairness among co-running applications (e.g. improving the fairness by up to a factor of 19.8 and 1.7 compared to the default barrier implementation of OpenMP in Completely Fair Scheduler and Balance Scheduling, respectively).