Hi all,
I test ping latency in two pass-through(KVM-VFIO) guests booted on
Debian 9.3 host(kernel version: 4.9.82, qemu:2.8.1).
The network card is 82599, and the two assigned VFs are on same port
eth0, config the ip for each VM, one is 192.168.0.1 and the other is
192.168.0.2. On 192.168.0.1 we run below command:
# ping 192.168.0.2 -i 0.01 outputfile_0530 &
After a night, I get several outliers from the ping results, almost
ping latency is 0.0xxms, but the outliers greater than 10ms, even up
to 1000ms.
I have see this phenomenon for several combination of host(Centos
7.3/Debian 9.3) and guest , it appears on Intel(R) Xeon(R) CPU E5-2680
v4 @ 2.40GHz, but Intel(R) Xeon(R) CPU E5-2670 v3 did not have this
issue.
By further analyze, I find it was associated with posted interrupt
feature. If disabled Interrupt Posting by adding "intremap=nopost" in
grub，this phenomenon disappeared, that is to say, all the ping latency
are litter than 10ms.
Additionally, add "idle=poll" in guest grub, it should send
interrupt w/o VM exit almostly, and it is proved by watching "PIW"
interrupt number unchanged. However, it becomes more likely to appear
latency bigger than 10ms.
Posted interrupt should decrease latency to my understand, but the
result is on the contrary. Is there something wrong with my test or
understand, Any advises?
Thanks!
Zhu Yijun
