On 2017-06-01 14:22, Per Jessen wrote:
Carlos E. R. wrote:
On 2017-06-01 08:31, Per Jessen wrote:
Some more examples -
on machines with uptimes between 30 and 60 days, depending on their load, I see drop rates of up to 1%. 0.99% on a xen guest, fairly busy.
On a mirror+squid, uptime 45days, I see 7% drop on one interface.
On four xen guests, not the most potent host machine, each has 10-11% dropped rx packages ...
two webservers, uptime 1600 days and 92 days, 0 rx drops.
The only possible pattern here might be the xen guests which exhibit somewhat higher rates than physical machines. My guess is that a machine dropping packages is simply busy - too much incoming traffic or too busy to process the interrupts.
Question.
The network "hardware" in the xen guest, is real or emulated?
They're virtual network devices, bridged with the physical device.
Because if it is emulated or virtual, not real hardware, it depends on the host CPU having time to attend it at the precise instant it is needed. A real network card will have some internal memory and processing power, and probably interrupts the machine only when the data chunk is ready. Perhaps even moves the chunk to main ram via dma, thus no cpu load.
The physical card still works the way it always has, but there is obvioulsy more code to be executed to get a network interrupt serviced in a guest. Something like that.
I have to wonder at how/if the card can write to ram that belongs to the guest. :-? :-o -- Cheers / Saludos, Carlos E. R. (from 42.2 x86_64 "Malachite" at Telcontar)