[ofiwg] libfabric fi_getinfo question

[Dave Goodell (dgoodell)]

Hi Vailiy,

This issue might be better discussed as a GitHub issue to make sure nobody loses track of it: https://github.com/ofiwg/libfabric/issues/new

But we can take a shot at it here on the list.

On Jul 16, 2015, at 4:14 AM, Vasiliy Tolstov <v.tolstov at selfip.ru> wrote:

> Hello, i'm try to understand usage of libfabric and stuc at simple
> example, it returns
> fi_getinfo(): info.c:34, ret=-22 (Invalid argument)
> 
> i'm try to resolve local node address with hostname =
> cn33.z1.mn2.simplecloud.ru and node_desc = cn33.z1.mn2.simplecloud.ru

What version of libfabric are you using?  The v1.0.0 release or something from git?

I don't have much experience with the verbs provider, but what does setting "FI_LOG_LEVEL=info" in your environment before running your test tell you?

If librdmacm/libibverbs is actually generating the underlying error (which seems likely), then "ltrace" might be helpful in figuring out what's going on.  If you let us know what sort of failure mode you're seeing then we can improve the logging output to make the failure clearer and/or fix any bugs that you reveal.

> P.S does it possible to resolve node address via libfabric that
> entered in node_desc
> (/sys/devices/pci0000:00/0000:00:02.0/0000:02:00.0/infiniband/mlx4_0/node_desc)

I don't know anything about this, sorry.

> int main(int argc, char **argv) {
>        int ret;
>        hints = fi_allocinfo();
>        if (hints == NULL) {
>          fprintf(stderr, "%s\n", "failed to alloc info");
>          return -1;
>        }
>        hints->addr_format = FI_SOCKADDR_IB;

I'm not too familiar with FI_SOCKADDR_IB (or AF_IB, which is what it maps to).  Have you previously used AF_IB successfully in your librdmacm applications on this system?

-Dave