kvm 性能调优

CPU Tuning

    网易运维团队测试得出：2 个 vcpu 分别绑定到不同 numa 节点的非超线程核上和分配到一对相邻的超线程核上的性能相差有 30%~40%(通过 SPEC CPU2006 工具测试)。可见，同一个虚拟机的 vcpu 需限定在同一个 NUMA 节点，并且分配该 NUMA 节点下的内存给虚拟机，保证虚拟机尽可能访问 local memory 而非 remote memory。

关于 memory allocation mode: If memory is overcommitted in strict mode and the guest does not have sufficient swap space, the kernel will kill some guest processes to retrieve additonal memory. Red Hat recommends using preferred allocation and specifying a single nodeset (for example, nodeset='0') to prevent this situation[1].

IRQ tuning

    CPU0 常用于处理中断请求，本身负荷较重[3]。事实上，线上环境 CPU0 完全处理 eth0 的 IRQ，eth1 的 IRQ 完全由另一个 CPU 处理，同时 CPU0 处理着大量的 CAL 类型中断。预留 2 个或者 4 个物理 CPU，这些 CPU 和对应网卡中断做绑定，使得其它 CPU 更好更完整的为云主机所用。  

VCPU topology tuning

Selecting any desired number of sockets, but with only a single core and a single thread usually gives the best performance results[1]. 即 VCPU 的 topology 设置为 sockets = vcpu_number, cores = 1, threads = 1. 

Disk IO Tuning

Disk IO cache Tuning

    kvm 支持多种虚拟机多种 IO Cache 方式：writeback, none, writethrough 等。性能上：writeback > none > writethrough，安全上 writeback < none < writethrough。For the best storage performance on guest operating systems that use raw disk volumes or partitions, completely avoid the page cache on the host[2].

                  

Disk IO scheduler

cfq 参数调优。

Memory Tuning

1. 关于 zone_reclaim_mode，设置为 disable 。

2. 关于 swappiness 参数，若 CPU macro-architecture 非 Intel Nehalem 架构，无需要修改优化。

3. 关闭 KVM 内存共享：When KSM merges across nodes on a NUMA host with multiple guest virtual machines, guests and CPUs from more distant nodes can suffer a significant increase of access latency to the merged KSM page[1].

Disable KSM 方式有两种： 

  1) 禁止某个 Guest 与其它 Guest 共享内存，XML 文件可配置为

  2)  禁止所有 Guest 直接共享内存，Host 配置为

4. 打开透明大页：KVM guests can be deployed with huge page memory support in order to improve performance by increasing CPU cache hits against the Transaction Lookaside Buffer (TLB)[1].  

打开透明大页方式有两种：

  1) 允许某个 Guest 开启透明大页

 2) 允许 Host 中所有 Guest 开启透明大页

Network IO Tuning

开启 vhost_net 模式[3]。 

 

参考资料

1. 1. Red Hat Enterprise Linux 6 Virtualization Tuning Optimization Guide
   2. Tuning KVM for performance    IBM
   3. 网易 openstack 部署运维实战