RE: [openib-general] MTHCA features
Title: RE: [openib-general] MTHCA features We recommend to work with FMRs and not memory windows due to performance. FMRS are much faster and available for kernel modules only. They are not yet implemented in mthca but it is possible to add them. Tziporet -Original Message- From: James Lentini [mailto:[EMAIL PROTECTED]] Sent: Tuesday, March 01, 2005 10:58 PM To: Roland Dreier Cc: openib-general@openib.org Subject: Re: [openib-general] MTHCA features NFS/RDMA doesn't require memory windows but it will make use of them if they are available. -james ___ openib-general mailing list openib-general@openib.org http://openib.org/mailman/listinfo/openib-general To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
Re: [openib-general] MTHCA features
NFS/RDMA doesn't require memory windows but it will make use of them if they are available. -james On Tue, 1 Mar 2005, Roland Dreier wrote: James> I ask because the DAT API contains interfaces that allow James> users to interact with memory windows. Are there any real applications that use those interfaces? Thanks, Roland ___ openib-general mailing list openib-general@openib.org http://openib.org/mailman/listinfo/openib-general To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
Re: [openib-general] MTHCA features
Or> By "performance characteristics" do you mean the extra Or> overhead to generate another rkey for the already registered Or> address range (and also to create/free the mw)? No, I mean the performance cost of binding/unbinding the MW. - R. ___ openib-general mailing list openib-general@openib.org http://openib.org/mailman/listinfo/openib-general To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
Re: [openib-general] MTHCA features
James> I ask because the DAT API contains interfaces that allow James> users to interact with memory windows. Are there any real applications that use those interfaces? Thanks, Roland ___ openib-general mailing list openib-general@openib.org http://openib.org/mailman/listinfo/openib-general To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
Re: [openib-general] MTHCA features
roland> James> It is my understanding that the current MTHCA driver does roland> James> not support InfiniBand memory windows or memory roland> James> registration using virtual addresses. roland> roland> James> Is this information correct? If so, when will these roland> James> features be supported? roland> roland> Well, memory registration is pretty complete. By design, we only roland> support memory registration with physical addresses for kernel roland> consumers even at the verbs API level (ie there are no mthca-specific roland> limitations). In the kernel, registration by virtual address is not roland> very useful. For userspace verbs, only registration by virtual roland> address is supported for obvious reasons. roland> roland> Memory windows are not implemented for mthca. It wouldn't be a lot of roland> work for someone with access to Mellanox documentation to implement roland> them, but they're not particularly useful due to their performance roland> characteristics. Is anyone on this list working on memory window support? I ask because the DAT API contains interfaces that allow users to interact with memory windows. ___ openib-general mailing list openib-general@openib.org http://openib.org/mailman/listinfo/openib-general To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
RE: [openib-general] MTHCA features
Roland > Memory windows are not implemented for mthca. It wouldn't be a lot of Roland > work for someone with access to Mellanox documentation to implement Roland > them, but they're not particularly useful due to their performance characteristics. By "performance characteristics" do you mean the extra overhead to generate another rkey for the already registered address range (and also to create/free the mw)? I guess an app that registers huge portions of memory in batch/startup time can also doing mw-ing in this time and not in the fast flow. It then have at hand pool of rkeys which it can use and recycle at run time. Or. ___ openib-general mailing list openib-general@openib.org http://openib.org/mailman/listinfo/openib-general To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
Re: [openib-general] MTHCA features
James> It is my understanding that the current MTHCA driver does James> not support InfiniBand memory windows or memory James> registration using virtual addresses. James> Is this information correct? If so, when will these James> features be supported? Well, memory registration is pretty complete. By design, we only support memory registration with physical addresses for kernel consumers even at the verbs API level (ie there are no mthca-specific limitations). In the kernel, registration by virtual address is not very useful. For userspace verbs, only registration by virtual address is supported for obvious reasons. Memory windows are not implemented for mthca. It wouldn't be a lot of work for someone with access to Mellanox documentation to implement them, but they're not particularly useful due to their performance characteristics. - R. ___ openib-general mailing list openib-general@openib.org http://openib.org/mailman/listinfo/openib-general To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
