Re: context switches and responsiveness

[Avi Kivity <avi@redhat.com>]

Antoine Martin wrote:
> -----BEGIN PGP SIGNED MESSAGE-----
> Hash: SHA512
>
> Hi,
>
> I've recently increased the number of guests (+20%) running on one of my
> hosts and found that the responsiveness suffered.
>   

How many guests?  How many vcpus?

> Before that, the context switches were hovering around 10K, now they're
> close to 30K. Or this could just be because I upgraded the host kernel
> and kvm: host and guests are running 2.6.29.4 + kvm-86
> Is there anything that can be done to reduce this?
> I thought dynticks would prevent unecessary context switching?
>   

dynticks is the default IIRC.  What type guests are you running?

If Linux, make sure the guests are dynticks as well.

> As you can see from the dstat sample below, the host can be very quiet
> at times, and yet the guests are a bit sluggish. (there is no swapping
> going on in the guests either)
> Is 30K context switches normal for a host that's mostly idle??
>   

Depends on guest count!  Please provide kvm_stat output.

> Here is a sample guest command line:
> - -clock dynticks -usbdevice tablet -m 1024 -monitor
> telnet:127.0.0.1:10005,server,nowait -L ./ -kernel ./bzImage-2.6.29.4
> - -append  earlyprintk=serial,ttyS0,115200,keep console=ttyS0,115200
>   

Ah, -usbdevice tablet, interrupt generator from hell.  Let it go and 
you'll see your context switch rate drop.

> - ----total-cpu-usage---- -dsk/total- -net/total- ---paging-- ---system--
> usr sys idl wai hiq siq| read  writ| recv  send|  in   out | int   csw
>  14  15  71   0   0   0|   0     0 | 942B 6912B|   0     0 |  20k   29k
>  11  17  71   0   0   0|   0   160k| 460B  818B|   0     0 |  19k   30k
>   

Interrupt rate is close to context switch rate, which is good.


-- 
error compiling committee.c: too many arguments to function