[mpich2-dev] mpiexec to call launcher for all names in machines file without hostname resolution (was -nolocal option)

[John Marshall]

On 07/09/2011 12:30 AM, Pavan Balaji wrote:
>
> On 07/08/2011 11:29 PM, John Marshall wrote:
>> On 07/09/2011 12:18 AM, Pavan Balaji wrote:
>>>
>>> On 07/08/2011 11:11 PM, John Marshall wrote:
>>>> Perhaps there is some confusion here, but I don't see how my launcher
>>>> could be called since *is_local == 1.
>>>
>>> I'm confused on what exactly you need. Do you want Hydra to not fall back to fork when a local hostname is provided? Why?
>> Right. I want my launcher to be called for each entry in the machines list, even if that entry ends up on the local host. No
>> fall back to fork, ever.
>>
>> Why? Because the entries in the machines file are _not_ necessarily hostnames. What they are is interpreted by the launcher
>> which effectively maps between the machines file entries and targets.
>
> I understand that they are not hostnames -- in such cases the host resolution will fail and you'll get this behavior anyway. 
> Are you not getting this behavior?
My point is that, for what I am doing, any sort of host resolution is problematic because it conflates the namespace of the 
machines file to be hostnames and not hostnames (i.e., when host resolution fails). I need for the machines file to be the 
namespace of not hostnames.

As a convoluted example, suppose a machine had multiple hostnames: name0, name1, name2, and localhost. If I ran mpiexec from 
that machine with the following machines file:

name0
name1
name2
localhost

as it is now and with your patch, hydra would fork, _not_ call my launcher. Instead, I want my launcher to be called each time, 
no forking. Why? Because for my situation name0, name1, name2, and localhost are not hostnames but labels that my launcher 
interprets.

John