[ofa-general] New proposal for memory management

[Matthew Koop]

I'm sure sure everyone has gotten this point yet -- It's not a matter of
the MPIs having this complicated memory registration code in userspace
they want to push into the kernel to simplify their lives.

The problem is that this registration cache in userspace is a hack that
can't even be guaranteed. Besides that, this hack ruins many memory
debugging tools.

This alone is a major hassle for many users since the memory registration
caching changes the timings. A user can't be told to just run the
application with the registration cache on for normal runs and then debug
with it off. Many errors end up being timing dependant.

Matt

On Fri, 1 May 2009, Woodruff, Robert J wrote:

> Jeff wrote,
>
> >It sounds like your main objection to fixing them is "it's too much
> >work."  :-(
>
> Not really, in general I think the kernel folks like to keep stuff
> out of the kernel if it is not really needed, i.e., if it can be implemented
> in user-space, especially really complicated
> things like this. It is probably a somewhat tricky code to implement, prone
> to bugs that could cause instability in the kernel. Remember if there is a bug
> in user-space and an application dies, it only effects that one application,
> if the bug is in the kernel and it crashes the system it affects everyone.
>
> That said, if you can find someone to implement it, then do it and send in
> the patches. Assuming the code is not too ugly, maybe it would get accepted.
>
>
> _______________________________________________
> general mailing list
> general at lists.openfabrics.org
> http://lists.openfabrics.org/cgi-bin/mailman/listinfo/general
>
> To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
>