UVA 4 tesla M2050 P2P problem

hi im hoping that I can find some answers here.
Im working with 4 tesla M2050 on “SuperServer 7046GT-TRF-FC475” server.
and im working with cuda C on Ubuntu 11.04 with CUDA5

when I query with cudaDeviceCanAccessPeer()
peer access between GPU0 <=> GPU1 and GPU0 <=> GPU1
but the access between GPU1 <=> GPU2 can not communicate
I think this is due to the pci-e controllers but i would like to know if someone could clarify this for me.

the reason i am asking is that i wanted to use data that resides one one GPU to do calculations on the rest

You can go through this document. http://docs.nvidia.com/cuda/gpudirect-rdma/index.html. I think this document will address your query.