On Mon, 2008-03-17 at 09:43 +0200, Or Gerlitz wrote: > Rajouri Jammu wrote: > > I was successfully able to register a total of 10GB of memory in 10M > > chunks. In other words 1024 MRs of 10MB each. > whatever. > > The thing is that a limitation --exists-- and my thinking was that > usage/registration of huge-pages should both allow the app to register > more memory and experience less cache misses (on HCAs where the IB > network MMU has an associated cache) since each slot used to translate > larger amount of memory. > > I would be glad to hear what others think on this point... >
Using a huge pages is an option and will benefit from what you mentioned. We used to have support for huge pages but it was not so trivial and there was no support from the kernel for such operations. I don't know what is the situation today. Another option but that's only for kernel consumers, is to use ib_reg_phys_mr where you can specify arbitrary physical regions sizes. _______________________________________________ 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
