[ofa-general] New proposal for memory management

Woodruff, Robert J robert.j.woodruff at intel.com
Thu Apr 30 13:45:33 PDT 2009


Jeff wrote,

>I'm a little amazed that it's gone this long without being fixed (I  
>know I spoke about this exact issue at Sonoma 3 years ago!).

If MPIs are so broken over OFA, then can you give me the list of applications
that are failing ? I did not hear from any customer at Sonoma that all
the MPIs are totally broken and do not work.
As far as I know, most if not all of the applications are running
just fine with the MPIs as they are today. At least I am not aware of
any applications that are failing when using Intel MPI. Sure, you need to hook
malloc and such and that is a bit tricky, but it seems like you
have figured out how to do it and it works.  If a notifier capability would make this easier, 
then perhaps that should be added, but adding a memory registration cache to the kernel,
that may or may not even meet the needs of all MPIs does not seem
like the right approach, it will just lead to kernel bloat.

Sure you have to manage your own cache, but you have chosen to
do a cache to get better performance. You did not have to do a cache at
all. You chose to do it to make your MPI better. 

To me, all this sounds like a lot of whining.... 
Why can't the OS fix all my problems.




More information about the general mailing list