[ofiwg] detecting FABRIC_DIRECT mismatch

Jason Gunthorpe jgunthorpe at obsidianresearch.com
Fri May 13 15:48:05 PDT 2016

On Fri, May 13, 2016 at 10:20:32PM +0000, Hefty, Sean wrote:
> provider may also replace certain constants and data structures with
> hardware specific versions.  This is how the mismatch is showing up.

That is still an ABI change, you can't have

 fi_foobar(struct fi_foo *)
 fi_foobar(struct fi_foo_verbs *)

With the same symbol name - that is an incorrect way to use dynamic
library symbols.

> Outside of select apps (e.g. benchmarks) running on exascale class
> machines, I doubt anyone will use FABRIC_DIRECT.

I don't know, I thought it was a great idea - most machines are
uniform, if I can gain a few percent just with a compiler option then
why not do it???

The key is to make it usable, which is what I'm talking here with
correct use of dynamic linking :)


More information about the ofiwg mailing list