• Christopher Yeoh's avatar
    Cross Memory Attach · fcf63409
    Christopher Yeoh authored
    The basic idea behind cross memory attach is to allow MPI programs doing
    intra-node communication to do a single copy of the message rather than a
    double copy of the message via shared memory.
    The following patch attempts to achieve this by allowing a destination
    process, given an address and size from a source process, to copy memory
    directly from the source process into its own address space via a system
    call.  There is also a symmetrical ability to copy from the current
    process's address space into a destination process's address space.
    - Use of /proc/pid/mem has been considered, but there are issues with
      using it:
      - Does not allow for specifying iovecs for both src and dest, assuming
        preadv or pwritev was implemented either the area read from or
      written to would need to be contiguous.
      - Currently mem_read allows only processes who are currently
      ptrace'ing the target and are still able to ptrace the target to read
      from the target. This check could possibly be moved to the open call,
      but its not clear exactly what race this restriction is stopping
      (reason  appears to have been lost)
      - Having to send the fd of /proc/self/mem via SCM_RIGHTS on unix
      domain socket is a bit ugly from a userspace point of view,
      especially when you may have hundreds if not (eventually) thousands
      of processes  that all need to do this with each other
      - Doesn't allow for some future use of the interface we would like to
      consider adding in the future (see below)
      - Interestingly reading from /proc/pid/mem currently actually
      involves two copies! (But this could be fixed pretty easily)
    As mentioned previously use of vmsplice instead was considered, but has
    problems.  Since you need the reader and writer working co-operatively if
    the pipe is not drained then you block.  Which requires some wrapping to
    do non blocking on the send side or polling on the receive.  In all to all
    communication it requires ordering otherwise you can deadlock.  And in the
    example of many MPI tasks writing to one MPI task vmsplice serialises the
    There are some cases of MPI collectives where even a single copy interface
    does not get us the performance gain we could.  For example in an
    MPI_Reduce rather than copy the data from the source we would like to
    instead use it directly in a mathops (say the reduce is doing a sum) as
    this would save us doing a copy.  We don't need to keep a copy of the data
    from the source.  I haven't implemented this, but I think this interface
    could in the future do all this through the use of the flags - eg could
    specify the math operation and type and the kernel rather than just
    copying the data would apply the specified operation between the source
    and destination and store it in the destination.
    Although we don't have a "second user" of the interface (though I've had
    some nibbles from people who may be interested in using it for intra
    process messaging which is not MPI).  This interface is something which
    hardware vendors are already doing for their custom drivers to implement
    fast local communication.  And so in addition to this being useful for
    OpenMPI it would mean the driver maintainers don't have to fix things up
    when the mm changes.
    There was some discussion about how much faster a true zero copy would
    go. Here's a link back to the email with some testing I did on that:
    There is a basic man page for the proposed interface here:
    This has been implemented for x86 and powerpc, other architecture should
    mainly (I think) just need to add syscall numbers for the process_vm_readv
    and process_vm_writev. There are 32 bit compatibility versions for
    64-bit kernels.
    For arch maintainers there are some simple tests to be able to quickly
    verify that the syscalls are working correctly here:
    http://ozlabs.org/~cyeoh/cma/cma-test-20110718.tgzSigned-off-by: default avatarChris Yeoh <yeohc@au1.ibm.com>
    Cc: Ingo Molnar <mingo@elte.hu>
    Cc: "H. Peter Anvin" <hpa@zytor.com>
    Cc: Thomas Gleixner <tglx@linutronix.de>
    Cc: Arnd Bergmann <arnd@arndb.de>
    Cc: Paul Mackerras <paulus@samba.org>
    Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
    Cc: David Howells <dhowells@redhat.com>
    Cc: James Morris <jmorris@namei.org>
    Cc: <linux-man@vger.kernel.org>
    Cc: <linux-arch@vger.kernel.org>
    Signed-off-by: default avatarAndrew Morton <akpm@linux-foundation.org>
    Signed-off-by: default avatarLinus Torvalds <torvalds@linux-foundation.org>
compat.h 19.4 KB