Is co-scheduling too expensive for SMP VMs?

Abstract

Symmetric multiprocessing (SMP) virtual machines (VMs) allow users to take advantage of a multiprocessor infrastructure. Despite the advantage, SMP VMs can cause synchronization latency to increase significantly, depending on task scheduling. In this paper, we show that even if a SMP VM runs non-concurrent applications, the synchronization latency problem can still occur due to synchronization in the VM kernel.Our experiments show that both of the widely used open source hypervisors, Xen and KVM, with the default schedulers are susceptible to the synchronization latency problem. To remediate this problem, previous works propose a co-scheduling solution where virtual CPUs (vCPUs) of a SMP VM are scheduled simultaneously. However, the co-scheduling approach can cause CPU fragmentation that reduces CPU utilization, priority inversion that degrades I/O performance, and execution delay, leading to deployment impediment. We propose a balance scheduling algorithm which simply balances vCPU siblings on different physical CPUs without forcing the vCPUs to be scheduled simultaneously. Balance scheduling can achieve similar or (up to 8%) better application performance than co-scheduling without the co-scheduling drawbacks, thereby benefiting various SMP VMs. The evaluation is thoroughly conducted against both concurrent and non-concurrent applications with CPU-bound, I/Obound, and network-bound workloads in KVM. For empirical comparison, we also implement the coscheduling algorithm on top of KVM's Completely Fair Scheduler (CFS). Compared to the synchronizationunaware CFS, balance scheduling can significantly improve application performance in a SMP VM (e.g. reduce the average TPC-W response time by up to 85%).

Towards fair and efficient SMP virtual machine scheduling

“…A possible solution is to balance the run queue weight (as shown in Figure 2(b)) to minimize vCPU weight dilution. However, perfect weight balance on different pCPUs is not always possible and such balancing may introduce the vCPU stacking issue [27], which limits the effective parallelism and increases the synchronization latency in SMP VMs. In production systems, VM management software often pre-calculates the fair share of individual VMs and applies a resource cap to each VM, preventing it from consuming excessive CPU time.…”

“…Synchronization overhead has been accused of limiting the scalability and performance of parallel programs in a virtualized environment [7,11,27,32]. Lock-holder preemption (LHP) [32] and vCPU stacking [27] are two uniques issues that arise with SMP VM scheduling.…”

“…Lock-holder preemption (LHP) [32] and vCPU stacking [27] are two uniques issues that arise with SMP VM scheduling. We show that mechanisms for enforcing fairness inadvertently exacerbate these synchronization issues.…”

“…PIN binds vCPUs to pCPUs and ensures that no vCPUs belonging to the same VM reside on one pCPU. It is similar to balance scheduling [27] and may incur imbalanced CPU allocation. Load-balancing (LB) allows vCPU migrations for fairness.…”

“…Literature has identified the synchronization overhead in SMP VMs as the major performance bottleneck for parallel programs in virtualized environments. Lock-holder preemption (LHP) [32] and vCPU stacking [27] are two unique issues arise with SMP scheduling. To mitigate these issues, researchers have proposed relaxed scheduling [29], balance scheduling [27], demand-based coordinated scheduling [11], and scheduling approaches based on spinlock-detection [7,9,32,34].…”