VM Scheduling Optimization For HPC Applications In Cloud

Cloud increasingly becomes a viable alternative to cluster for high performance parallel applications, especially for many small and medium scale users. However, when running such an applications in cloud platforms, virtual machines(VMs) are easily affected by synchronization latency problem. The problem leads to a serious performance degradation. There are two main reasons, the first is that a host is unaware of the synchronization requirements of its guests. The second is the synchronization requests of a VM on one node are agnostic to the VM on the other nodes. It's a huge challenge to mitigate the negative influence of virtualization on synchronization and accelerate the synchronization responses of HPC applications.In order to solve the two synchronization issues mentioned above, a two phases synchronization-aware(TPS) scheduling algorithm is proposed. This algorithm considers the synchronization demands both intra-VM and inter-VM. It uses spinlocks and network packets as the metrics to detect the synchronization demands and then schedules VMs based on spinlock-aware and communication-aware strategies. We implement the algorithm based on KVM and evaluate the experiment in our cluster environment. Experiments show that our TPS strategy obtains a better performance, about 30%~40%, for HPC applications and avoids the negative impact on non-HPC applications. What's more, TPS scheduling shows a similar fairness with CFS scheduling. Therefore, our approach is an effective solution to the HPC synchronization issue in cloud.