> On 9 Oct 2018, at 11:03, Håkon Bugge <haakon.bugge@xxxxxxxxxx> wrote: > > > >> On 8 Oct 2018, at 15:07, Hal Rosenstock <hal@xxxxxxxxxxxxxxxxxx> wrote: >> >> Hi, >> >> On 10/8/2018 3:32 AM, Håkon Bugge wrote: >>> Hi, >>> >>> >>> ibacm uses the triple Node GUID, Port number, and pkey to designate an endpoint. It is currently a hard-limit of four addresses that can be associated with an endpoint. >>> >>> Now, this works well for most deployments. >>> >>> In Oracle, we have some tests where we enable 16 VFs on an HCA (actually, any number), and use these VFs in a bare-metal fashion, i.e., we have 34 IPoIB devices and no VMs. >>> >>> In this case, the above ibacm limitation hits us. This because we will have 17 IP addresses associated with the triple Node GUID, Port #, and pkey. The Node GUID is the very same for all the VFs. >>> >>> Would it make better sense for ibacm to use the tuple Port GUID and pkey to designate and endpoint instead? Well, just chatted with Bjørn Dag (BD) here. He mentioned that the reason for choosing the Node GUID could be suport for APM. Is that so? Thxs, Håkon >> >> Makes sense to me since all endports are required to have port GUID. >> Only issue I see is backward compatibility (supporting old address >> config files). > > If that is an issue, may be ib_acme can be used to transform the address file. > > Right now I am working on some other bugs that have priority. > > > Thxs, Håkon > >> >> -- Hal >> >>> >>> Thxs, Håkon
