Christopher Porter wrote:
> Jeff -
>
> Thanks for the direct response.
>
> 1) The practice of disallowing rsh/ssh to all computational hosts is the "stick" used in several corporate environments to make sure users are submitting their jobs to the workload management system [WLM] (LSF in this case) instead of running directly on the systems. In an environment with 1000s of computational hosts and 1000s of users where it's very important to know who ran what for how long, getting people to use the WLM is important.
All of which can be done by having job submissions handled by a gateway,
and then leveraging the considerable power of ssh at the backend, without
expecting the mpi implementation to encapsulate arbitrary queue management
systems. This is a task best suited to a modular approach which makes
use of existing and mature interfaces.
I will drop out of this now discussion since I can't seem to get my
head around the management driven nature of the problem.
|
