If it ain't broke, don't fix it. I am more than skeptical about the
interest of this new notation.
The two behaviors you describe for include and exclude do not look
conflicting to me. Inclusion is a strong request, the user enforce the
usage of a specific interface. If the interface is not available, then
we have a problem. Exclude on the other side, must enforce that a
specific interface is not in use, fact that can be quite simple if the
interface is not available.
I'm not a fan of the nowarn option. Seems like a lot of code with
limited interest, especially if we only plan to support it in TCP.
If you need specialized arguments for some of your nodes here is what
I do: rename the binaries to .orig, and use the original name to
create a sh script that will change the value of mca_param_files to
something based on the host name (if such a file exists) and then call
the .orig executable. Works like a charm., even when a batch scheduler
On Mon, Feb 4, 2013 at 12:02 PM, Jeff Squyres (jsquyres)
> On Feb 1, 2013, at 9:59 PM, "Barrett, Brian W" <bwbarre_at_[hidden]> wrote:
>> I don't think this is right either. Excluding a device that doesn't exist has many use cases. Such as disabling a network that only exists on part of the cluster. I'm not sure about what to do with seq; it's more like include than exclude.
> Hmm. I've now given this quite a bit of thought. Here's what I think:
> 1. Just like there might be good reasons to exclude non-existent interfaces (e.g., networks that only include on part of the cluster), the same argument could be made for *including* non-existent interfaces.
> 2. It seems odd to me to have different behavior for non-existent interfaces between include, exclude, and/or seq.
> 3. We have a very strong precedent throughout OMPI that if a human asks for something that OMPI can't deliver, OMPI should error. According to this, and according to the Law of Least Surprise, I would think that if I typo an exclude interface name, OMPI should error and make a human figure it out.
> 4. If someone wants different includes/excludes in different parts of the cluster, then they should have per-node values for these MCA params.
> 5. That being said, #4 is not always feasible. Concrete example (which is why this whole thing started, incidentally): in my MTT cluster at Cisco, I have *some* nodes with back-to-back interfaces. I can't think of a good way to have per-node MCA params in an MTT run that is SLURM-queued and may end up on random nodes in my cluster -- that may or may not include nodes with loopback interfaces.
> So how about this compromise:
> If an invalid include, exclude, or if_seq interface is specified:
> - If that interface is prefaced with "nowarn:", silently ignore that token
> - Otherwise, display a show_help message and ignore the TCP BTL
> For example:
> mpirun --mca btl_tcp_if_include nowarn:eth5,eth6
> - If eth5 doesn't exist, the job will continue just as if eth5 wasn't specified
> - If eth6 doesn't exist, the TCP BTL will disqualify itself
> (BTW: yes, I'm volunteering to code up whatever we agree on)
> Jeff Squyres
> For corporate legal information go to: http://www.cisco.com/web/about/doing_business/legal/cri/
> devel mailing list