Description of problem:
During a testing of CPU Cgroup, we observe load is not balanced among multiple cores.

Version-Release number of selected component (if applicable):
RHEL6.1 with kernel-2.6.32-131.12.1

How reproducible:
code 1: a CPU busy loop code, called bbb.c:
int main()
{
    while(1)
    {
    };
}
code 2: a bash script called test.sh:
#!/bin/sh
count=0
pids=" "
while [ $count -lt 32 ]
do
        mkdir /cgroup/$count
        echo 1024 > /cgroup/$count/cpu.shares
        ./bbb &
        pid=`echo $!`
        echo $pid > /cgroup/$count/tasks
        pids=`echo $pids" "$pid`
        count=`expr $count + 1`
done
echo "for pid in $pids;do cat /proc/$pid/sched|grep sum_exec_runtime;done" > show.sh
watch -n1 sh show.sh


Steps to Reproduce:
1. mount -t cgroup -o cpu cpuctl /cgroup
2. run script test.sh
3. observe processes by "top -d 1"
  
Actual results:
Some process may ocuupy 99% CPU time while some others only have 33%. Even we observe 2 or 3 processes occupy more than 90% CPU time all the time.
In long term, we see CPU time slices are averaged between all groups, but we observe processes are scheduled among different CPU cores to achive cpu-time-console. Which result response delay in my testing environment.

Expected results:
All groups should share CPU times. e.g. running 32 while-1-looping processes on 16 core machinese, we should see each process occupies ~50% CPU time, and no too much process migration between different cores.


Additional info:
1, for non-NUMA machinese, I don't observe the above issue. The issue can be easily reproduced on NUMA x86_64 machine.
2, I also tried the test case on vanilla kernel 3.1.0-rc9, the problem is *FIXED* in upstream kernel.
3, Now we are trying to figure out which patch solve the problem.