The cpu-partitioning profile should disable the kvmclock sync feature and the ple, just as the real-time host profile does. Otherwise, a process running on isolated CPUs will be interrupted by timers or unwanted VM exists.

Here's the code to do this (from /usr/lib/tuned/realtime-virtual-host/script.sh):

create_modfile()
{
    modinfo -p kvm | grep -q kvmclock_periodic_sync
    if [ "$?" -eq 0 ]; then
        echo "options kvm kvmclock_periodic_sync=0" > $modfile
    fi

    modinfo -p kvm_intel | grep -q ple_gap
    if [ "$?" -eq 0 ]; then
        echo "options kvm_intel ple_gap=0" >> $modfile
    fi
}

start()
{
    [...]

    if [ ! -f $modfile ]; then
        create_modfile
    fi

    [...]
}

stop()
{
    [...]
    [ "$1" = "profile_switch" ] && rm -f $modfile
    [...]
}

A few observations:

1. In the RT profiles, this code is only part of the host profile. But I guess there's no problem running it in the guest too (meaning, we can keep a single cpu-partitioning profile instead of breaking it into host and guest profiles)

2. modfile should probably be /etc/modprobe.d/kvm-cpu-partitioning.conf

3. Maybe we should create functions in /usr/lib/tuned/functions to handle /etc/modprobe.d files? So that we avoid code duplication like this?