Gleaner: Mitigating the blocked-waiter wakeup problem for virtualized multicore applications

As the number of cores in a multicore node increases in accordance with Moore's law, the question arises as to what are the costs of virtualized environments when scaling applications to take advantage of larger core counts. While a widely-known cost due to preempted spinlock holders has been extensively studied, this paper studies another cost, which has received little attention. The cost is caused by the intervention from the VMM during synchronization-induced idling in the application, guest OS, or supporting libraries-we call this the blocked-waiter wakeup (BWW) problem. The paper systematically analyzes the cause of the BWW problem and studies its performance issues, including increased execution times, reduced system throughput, and performance unpredictability. To deal with these issues, the paper proposes a solution, Gleaner, which integrates idling operations and imbalanced scheduling as a mitigation to this problem. We show how Gleaner can be implemented without intrusive modification to the guest OS. Extensive experiments show that Gleaner can effectively reduce the virtualization cost incurred by blocking synchronization and improve the performance of individual applications by 16x and system throughput by 3x.