Thanks for all the responses. If I have to register large amounts of memory from user space, it's still unclear to me what's the best (predictable) way to do it today?
Since there is a limit, how do I figure out the exact limit? As I was able to register almost 20GB in 10MB chunks I'm not sure why I didn't hit the 2^20*4K limit on the MTTs. Any clarifications would be most helpful. thanks! On Mon, Mar 17, 2008 at 3:12 AM, Jack Morgenstein <[EMAIL PROTECTED]> wrote: > On Monday 17 March 2008 09:22, Or Gerlitz wrote: > > Rajouri Jammu wrote: > > > Does that mean that no matter how I size my memory regions the maximum > > > amount of (total) memory I can register is = 2^20 * 4K = 4GB? > > > I.e., Am I limited by the total number of MTTs? > > Generally speaking, yes, the IB network MMU (MTT this case) is limited > > in the number of slots it has and so in the amount of memory it can map > > at once, but I assume this is typical eg for I/O MMUs and makes sense. > > > > However, each slot can map --more-- then 4K, so one should be able to > > use huge-pages etc. I am not sure what is the actual status of > > registering huge-pages by the Linux IB stack, maybe Roland can comment > > on that. > > > > Or. > > In Mellanox HCAs, each memory region (i.e., mpt) can specify how many > pages > each mtt entry it uses represents. This would allow a much larger number > of > large regions to be mapped (essentially eliminating the current > limitation). > > Currently, this feature is not supported by the driver code. > > - Jack > > > > > > >
_______________________________________________ general mailing list [email protected] http://lists.openfabrics.org/cgi-bin/mailman/listinfo/general To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
