ParMETIS_V3_PartMeshKway gives that most of the elements belong to another processor

Hi,
I'm trying to use ParMETIS_V3_PartMeshKway to compute a repartitioning for a distributed finite element mesh that has been adaptively refined locally on each processor. However, whatever I do, this routine gives as a result that most of the elements should belong to another processor than they currently reside on. Below is a small 3-processor case that illustrates the problem.

Input on processor 1:

elmdist: 1 19 32 51

eptr: 1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 52 55

eind: 2 12 1 12 2 13 3 13 2 13 3 14 4 14 3 14 4 15 4 5 36 15 4 36 15 36 35 36 5 34 35 34 16 34 35 36 13 23 12 23 13 24 14 24 13 24 14 25 15 25 14 25 15 26

Input on processor 2:

elmdist: 1 19 32 51

eptr: 1 4 7 10 13 16 19 22 25 28 31 34 37 40

eind: 9 19 8 19 9 20 10 20 9 20 10 21 11 21 10 21 11 22 29 19 30 20 30 19 30 20 31 21 31 20 31 21 32 22 32 21 32 22 33

Input on processor 3:

elmdist: 1 19 32 51

eptr: 1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 52 55 58

eind: 5 6 37 5 37 34 37 16 34 17 16 37 6 17 37 7 17 6 17 7 18 8 18 7 18 8 19 26 15 38 38 15 35 16 38 35 27 26 38 16 27 38 17 27 16 27 17 28 18 28 17 28 18 29 19 29 18

For all 3 processors, I use the parameters:

wgtflag=0
num_flag=1
ncon=1
ncommonnodes=2
nparts=3
tpwgts=3*0.333333
ubvec=1.050000
options=0 0 0

This gives me the result:

On processor 1:
part: 2 2 2 2 2 2 2 2 3 2 3 3 2 2 2 2 2 3
edgecut=7

On processor 2:
part: 1 1 1 1 1 1 1 1 1 1 1 1 1
edgecut=7

On processor 3:
part: 3 3 3 3 3 3 1 1 1 3 3 3 3 3 3 3 3 3 1
edgecut=7

So, it wants nearly all the elements on processor 1 to move to processor 2, in exchange of all elements currently on processor 2. But wouldn't it be better to let all elements currently on processor 2 stay there, only adding some extra ones from processor 1 (or 3) instead? Only for processor 3 the results are somewhat as expected.

Can someone explain why this happens, or what I am doing wrong?

Thanks,
Knut