[libfabric-users] Binding EP with memory region with flags?

Sur, Sayantan sayantan.sur at intel.com
Thu Jul 14 12:41:50 PDT 2016


Thanks for letting us know! If you do use it make sure you disable the SEGMENT_FAST and SEGMENT_LARGE code paths, due to the bug that I mentioned.

We’ll get that fixed soon.

-----Original Message-----
From: "Lim, Jintack" <jin-tack.lim at hpe.com>
Date: Thursday, July 14, 2016 at 12:18 PM
To: "Dave Goodell (dgoodell)" <dgoodell at cisco.com>, Sayantan Sur <sayantan.sur at intel.com>
Cc: "libfabric-users at lists.openfabrics.org" <libfabric-users at lists.openfabrics.org>
Subject: RE: [libfabric-users] Binding EP with memory region with flags?

    Thanks, Sayantan and Dave.
    
    In fact, I re-compiled GASNet without binding mr to ep, and it just worked.
    
    Thanks,
    Jintack
    
    -----Original Message-----
    From: Dave Goodell (dgoodell) [mailto:dgoodell at cisco.com]
    Sent: Thursday, July 14, 2016 11:36 AM
    To: Sur, Sayantan <sayantan.sur at intel.com>
    Cc: Lim, Jintack <jin-tack.lim at hpe.com>; libfabric-users at lists.openfabrics.org
    Subject: Re: [libfabric-users] Binding EP with memory region with flags?
    
    On Jul 14, 2016, at 1:31 PM, Sur, Sayantan <sayantan.sur at intel.com> wrote:
    >
    > I think we need to revisit that and fix the bug. In the meanwhile, ofi_ep_bind_valid() seems to be correct per the current definition of OFI … there shouldn’t be a need to bind a MR to an EP.
    
    Agreed.  fi_mr_reg() registers the memory with the domain, not with an endpoint.
    
    The man page for fi_endpoint(3) could probably use a tweak to avoid misleading users into thinking that binding MRs to EPs is necessary:
    
    ----8<----
           fi_ep_bind
                  Associate  an  endpoint  with  an event queue, completion queue,
                  counter, address vector, or memory region
    ----8<----
    
    -Dave
    
    



More information about the Libfabric-users mailing list