Re: [OMPI users] RDMA over IB between heterogenous processors with different endianness
In the entire 1.2 series, RDMA is only allowed if the architecture of the two processes match. The 1.3 series added the ability to chose based on datatype. Brian On Mon, 25 Aug 2008, Mi Yan wrote: Brian, I'm using OpenMPI 1.2.6 (r17946). Could you plese check which version works ? Thanks a lot, Mi Inactive hide details for "Brian W. Barrett" <brbar...@open-mpi.org> "Brian W. Barrett" <brbar...@open-mpi.org> "Brian W. Barrett" <brbar...@open-mpi.org> Sent by: users-boun...@open-mpi.org 08/25/2008 01:44 PM Please respond to Open MPI Users <us...@open-mpi.org> [IMAGE] To [IMAGE] Open MPI Users <us...@open-mpi.org> [IMAGE] cc [IMAGE] Greg Rodgers/Poughkeepsie/IBM@IBMUS, Brad Benton/Austin/IBM@IBMUS [IMAGE] Subject [IMAGE] Re: [OMPI users] RDMA over IB between heterogenous processors with different endianness [IMAGE][IMAGE] On Mon, 25 Aug 2008, Mi Yan wrote: > Does OpenMPI always use SEND/RECV protocol between heterogeneous > processors with different endianness? > > I tried btl_openib_flags to be 2 , 4 and 6 respectively to allowe RDMA, > but the bandwidth between the two heterogeneous nodes is slow, same as > the bandwidth when btl_openib_flags to be 1. Seems to me SEND/RECV is > always used no matter btl_openib_flags is. Can I force OpenMPI to use > RDMA between x86 and PPC? I only transfer MPI_BYTE, so we do not need the > support for endianness. Which version of Open MPI are you using? In recent versions (I don't remember exactly when the change occured, unfortuantely), the decision between send/recv and rdma was moved from being solely based on the architecture of the remote process to being based on the architecture and datatype. It's possible this has been broken again, but there defintiely was some window (possibly only on the development trunk) when that worked correctly. Brian ___ users mailing list us...@open-mpi.org http://www.open-mpi.org/mailman/listinfo.cgi/users
Re: [OMPI users] RDMA over IB between heterogenous processors with different endianness
I believe that this is something Brad at IBM worked on, wasn't it? I *think* you may just need the development trunk (i.e., upcoming v1.3), but I won't swear to that. Regardless, you need to have OMPI compiled for heterogeneous support because control headers still need to be adjusted for endian-ness, etc. (even if you're only sending MPI_BYTE). On Aug 25, 2008, at 1:57 PM, Mi Yan wrote: Brian, I'm using OpenMPI 1.2.6 (r17946). Could you plese check which version works ? Thanks a lot, Mi "Brian W. Barrett" <brbar...@open-mpi.org> "Brian W. Barrett" <brbar...@open-mpi.org> Sent by: users-boun...@open-mpi.org 08/25/2008 01:44 PM Please respond to Open MPI Users <us...@open-mpi.org> To Open MPI Users <us...@open-mpi.org> cc Greg Rodgers/Poughkeepsie/IBM@IBMUS, Brad Benton/Austin/IBM@IBMUS Subject Re: [OMPI users] RDMA over IB between heterogenous processors with different endianness On Mon, 25 Aug 2008, Mi Yan wrote: > Does OpenMPI always use SEND/RECV protocol between heterogeneous > processors with different endianness? > > I tried btl_openib_flags to be 2 , 4 and 6 respectively to allowe RDMA, > but the bandwidth between the two heterogeneous nodes is slow, same as > the bandwidth when btl_openib_flags to be 1. Seems to me SEND/RECV is > always used no matter btl_openib_flags is. Can I force OpenMPI to use > RDMA between x86 and PPC? I only transfer MPI_BYTE, so we do not need the > support for endianness. Which version of Open MPI are you using? In recent versions (I don't remember exactly when the change occured, unfortuantely), the decision between send/recv and rdma was moved from being solely based on the architecture of the remote process to being based on the architecture and datatype. It's possible this has been broken again, but there defintiely was some window (possibly only on the development trunk) when that worked correctly. Brian ___ users mailing list us...@open-mpi.org http://www.open-mpi.org/mailman/listinfo.cgi/users ___ users mailing list us...@open-mpi.org http://www.open-mpi.org/mailman/listinfo.cgi/users -- Jeff Squyres Cisco Systems
Re: [OMPI users] RDMA over IB between heterogenous processors with different endianness
Brian, I'm using OpenMPI 1.2.6 (r17946). Could you plese check which version works ? Thanks a lot, Mi "Brian W. Barrett" <brbarret@open-mp To i.org>Open MPI Users <us...@open-mpi.org> Sent by: cc users-bounces@ope Greg n-mpi.org Rodgers/Poughkeepsie/IBM@IBMUS, Brad Benton/Austin/IBM@IBMUS Subject 08/25/2008 01:44 Re: [OMPI users] RDMA over IB PM between heterogenous processors with different endianness Please respond to Open MPI Users <users@open-mpi.o rg> On Mon, 25 Aug 2008, Mi Yan wrote: > Does OpenMPI always use SEND/RECV protocol between heterogeneous > processors with different endianness? > > I tried btl_openib_flags to be 2 , 4 and 6 respectively to allowe RDMA, > but the bandwidth between the two heterogeneous nodes is slow, same as > the bandwidth when btl_openib_flags to be 1. Seems to me SEND/RECV is > always used no matter btl_openib_flags is. Can I force OpenMPI to use > RDMA between x86 and PPC? I only transfer MPI_BYTE, so we do not need the > support for endianness. Which version of Open MPI are you using? In recent versions (I don't remember exactly when the change occured, unfortuantely), the decision between send/recv and rdma was moved from being solely based on the architecture of the remote process to being based on the architecture and datatype. It's possible this has been broken again, but there defintiely was some window (possibly only on the development trunk) when that worked correctly. Brian ___ users mailing list us...@open-mpi.org http://www.open-mpi.org/mailman/listinfo.cgi/users
Re: [OMPI users] RDMA over IB between heterogenous processors with different endianness
On Mon, 25 Aug 2008, Mi Yan wrote: Does OpenMPI always use SEND/RECV protocol between heterogeneous processors with different endianness? I tried btl_openib_flags to be 2 , 4 and 6 respectively to allowe RDMA, but the bandwidth between the two heterogeneous nodes is slow, same as the bandwidth when btl_openib_flags to be 1. Seems to me SEND/RECV is always used no matter btl_openib_flags is. Can I force OpenMPI to use RDMA between x86 and PPC? I only transfer MPI_BYTE, so we do not need the support for endianness. Which version of Open MPI are you using? In recent versions (I don't remember exactly when the change occured, unfortuantely), the decision between send/recv and rdma was moved from being solely based on the architecture of the remote process to being based on the architecture and datatype. It's possible this has been broken again, but there defintiely was some window (possibly only on the development trunk) when that worked correctly. Brian