[libfabric-users] 2 questions

Biddiscombe, John A. biddisco at cscs.ch
Tue Feb 21 10:18:09 PST 2017

Sean, Dave

No bugs. I made the mistake of looking at examples and copying/pasting bits of code without properly reading the actual docs.

I had this idea in my mind that getinfo was to -get- information and I was expecting the fi_connect call to take the dest stuff and the info to just be for my enlightenment rather that a way of initiating the setup for the connection. As Dave pointed out – had I actually read things properly, I’d have seen the ‘service’ and ‘node’ params were there and I could use them too. I did not expect that for N connections, I’d need to call get_info N times, I thought it was a once only call (and I used it for the passive endpoint already).

I’ll get the hang of it soon. (hopefully)



On 21.02.17, 18:19, "Hefty, Sean" <sean.hefty at intel.com> wrote:

    > > To answer Q1 : I discover that you have to put the dest address into
    > the info->dest_addr fields before calling getinfo - which is a bit
    > annoying, but after stepping though the libfabric code, I see it is
    > using the rdma_cm functions under the hood and the rdma_get_peer_addr
    > fails if it hasn’t been set.
    > > (Note to self: find out how these address vectors work cos they look
    > like they will be much easier to work with)
    > You could set hints->dest_addr, but I think the more typical model is
    > to set the "node" and "service" parameters to fi_getinfo() to be
    > something like "" and "37000" on the active side, which
    > should set the returned info->dest_addr correctly for you.  I.e., it
    > should behave much like getaddrinfo(3).
    This is more typical.
    You shouldn't *need* to specify the destination address on the active side as part of getinfo, but if you don't, then you will need to ensure that you select an appropriate source address capable of communicating with the destination.
    It is possible that you're hitting a bug in the verbs provider handling this case.  If you can provide more details on your test, or change the pingpong test to behave similarly, we can investigate.
    - Sean 

More information about the Libfabric-users mailing list