SMP Architecture Server Performance Optimization Under Virtualization Environment

Virtualization technology, because of its management for resources and good isolation,has been widely used in test development environment, while some enterprises applicatevirtualization technology to contribute the excess out of server resources to provide userswith cloud services. The future of virtualization technology will be more popular in enter-prises, not only to provide users with better services, but also convenient for the allocationof enterprise resources management.This study analyzes server performance under SMP architecture of virtualized envi-ronment compared to existing non-virtualized environments. Distribution of the more un-stable and longer server response time on a virtualized server, deep researching for differ-ent binding strategies performance have been done, and found the server’s response timebecame more stable and shorter, while binding the Driver Domain and Guest Domain’svirtual CPUs to physical CPUs. Howerver, the binding of virtual CPUs and physical CPUsleading to NIC hardware interrupts and software interrupts not loading balancing on vir-tual CPUs, and directly resulting in not load balancing on the physical CPUs. Some CPUsthat handle large interrupts become the server performance bottleneck, resulting in the to-tal processing time of connections much longer. The paper presents the case of binding theDriver Domain and Guest Domain’s virtual CPUs to physical CPUs, balanced the CPUload by balancing the hardware interrupts and software interrupts. This method can avoidthe physical CPU becoming the bottleneck for server performance on SMP architecture invirtualization environment, while achieving the stability of server response time. Addi-tional experiment on SMP architecture in virtualized environments, we found Virtual Ma-chine Manager’s scheduling for virtual CPUs exist a loophole, resulting in virtual CPU’sscheduling interval too small when a large number of concurrent Network I/O arrives, thisbrought much high overhead. To end this, we present a schedule rate controller monitoringthe interval time between two virtual CPU’s scheduling. When the scheduling interval isless than the set threshold time, the schedule rate controller needs to determine whetherthis scheduling is necessary or not. This schedule rate controller can avoid meaninglessscheduling, and reduce the overhead of frequent scheduling, and finally improve the serv-er performance on SMP architecture in virtualization environment.Experiments show that the improved load balancing scheduling method can effec-tively improve server performance in virtualization, while achieving physical CPUs loadbalancing.