Open DyingOwl opened 1 year ago
Hi,
Never heard of CPU masks before, only CPU pinning. I'm not sure about the difference with masks :thinking:
Have you tried pinning? See https://support.citrix.com/article/CTX246466/how-to-pin-citrix-hypervisor-virtual-cpus-to-specific-physical-cpus
Hi,
Never heard of CPU masks before, only CPU pinning. I'm not sure about the difference with masks 🤔
Have you tried pinning? See https://support.citrix.com/article/CTX246466/how-to-pin-citrix-hypervisor-virtual-cpus-to-specific-physical-cpus
In my understanding, these are the same things (sometimes called NUMA affinity, either). I use this name in the title because the function needs to be configured through the VCPU mask parameter in the xe command. Refer doc
Hi~ I am trying to migrate from XEN (original official version) to XCP-ng. When I tried to configure the virtual machine, I found it difficult to ensure its real-time performance. In my previous XEN environment, the most effective method is to use
cpus = ['4','5']
parameter (for example) in xl.cfg file, which can bind vcpu0 and pcpu4, vcpu1 and pcpu5 of this VM. It can be confirmed in the following ways:With some other configurations in VM's grub and kernel, it can ensure that the maximum value measured by
cyclictest -p 80 -t5 -n -i 1000 -l ...
does not exceed 300us. In contrast, I found commandxe vm-param-set uuid=[VM-UUID] VCPUs-params:mask='4;5'
in XCP-ng, which can cause the same result as the previous. Before executingvm-param-set
, the maximum test result ofcyclictest
is about 4ms, but it exceeds 15ms (sometimes even 25ms) after execution and reboot. According to historical experience, providing VNC functions for VM may also affect test results. So when I log in through SSH(instead of XO or XCP-ng Center), the maximum can drop to 10ms. After I temporarily disable the vga of the VM by modifying file/usr/lib64/xen/bin/qemu-wrapper
, the maximum can drop to 2ms, but there is still a big gap with my previous environment.In the XCP-ng environment, is
VCPU mask
designed to optimize the stability of VCPU scheduling and the real-time performance of the virtual machine? If not, how can I ensure that the virtual machine has high real-time performance in the XCP-ng environment?