[libfabric-users] Failure of fi_getinfo with verbs backend

Jörn Schumacher joern.schumacher at cern.ch
Thu Jun 23 07:39:47 PDT 2016


I am trying to run a libfabric program with the verbs provider on a new 
system. The program works on a local test machine, but on the new system 
fi_getinfo fails. When I run fi_info on the new machine, the verbs 
provider is there and seems to have the same configuration as on my 
local testbed. The only difference is that addr_format on the new 
machine is FI_SOCKADDR_IN, whereas locally it is FI_SOCKADDR_IN6, but I 
use IPv4 addressing anyway.

The output of fi_info is here as well as a small test program that 
produces the error: 

Here is the relevant code snippet:
> hints = fi_allocinfo();
> hints->ep_attr->type  = FI_EP_MSG;
> hints->caps = FI_MSG;
> hints->mode   = FI_LOCAL_MR;
> if(ret = fi_getinfo(FI_VERSION(1, 1), "", "12345", 0, hints, &fi))
> {
>   ERROR("fi_getinfo failed: %d '%s'", ret, fi_strerror(-ret));
> }

Which fails with
> error: fi_getinfo failed: -61 'No data available'

The output of fi_info suggests the verbs provider has the capabilities I 
request. Oddly enough, it works on the server side, where the only 
difference is an added FI_SOURCE flag to the fi_getinfo call.

What could be a reason that fi_getinfo fails?

Thanks a lot,

More information about the Libfabric-users mailing list