On Mon, 2018-12-03 at 17:14 +0100, Håkon Bugge wrote:
> Thanks for the review, Doug!
You're welcome.
> I do not like this series either. Let me suggest; 1) I clean the
> commits up and 2), I go back to re-evaluate the use of Port GUID
> instead of Node GUID.
>
> The reason for 2) is that I do not think the ibacm address file will change for most cases, which I presume is a single HCA with one of two ports, and only one address assigned per port.
I wouldn't assume one address per port. Every P_Key on a port creates
additional addresses to be stored.
> After all, the address file doesn't contain any GUIDs. If the address file doesn't change, my take is to abandon this series and use Port GUIDs instead. The fix, in case of incompatibility, would simply be to delete the address file before starting ibacm.
>
> And me confused, since ibacm creates the address file if it is not
> there, why isn't created _every_ time ibacm is started? And if so, why
> is it needed at all?
For performance reasons I think. Doesn't it also store remote addresses
that it has learned over time? In a 10,000 node cluster, being able to
prime your machine name -> address mapping table instead of having to
relearn it on reboot is a big time saver. But I could be off, it's been
a while since I looked into ibacm.
--
Doug Ledford <dledford@xxxxxxxxxx>
GPG KeyID: B826A3330E572FDD
Key fingerprint = AE6B 1BDA 122B 23B4 265B 1274 B826 A333 0E57 2FDD
Attachment:
signature.asc
Description: This is a digitally signed message part
