Ian, Thank you so much for taking the time to write such a detailed answer. And thanks everyone who pitched in too. It's clear now.
Best regards, Leo. On Wed, Aug 1, 2018 at 8:22 PM Ian Buckley <ian.buck...@gmail.com> wrote: > Leo, > Ethernet does not fragment IP packets. There is a 1:1 relationship between > ethernet packet and IP packet. > > IP Fragmentation is very much a legacy concept, which in a modern network > adds complications we could do without. As early as the 1980’s these > limitations were being realized ( > http://www.hpl.hp.com/techreports/Compaq-DEC/WRL-87-3.pdf). IPv6 has done > away with it all together, enforcing a minimum MTU and requiring the use of > MTU path discovery to send larger packets. When I wrote the CHDR spec bits > were allocated as seemed best to give it longevity and flexibility, and > there were enough header bits that size could have a nice 2^n 16bit field. > At lower bitrates packets this large would have latencies that would likely > make them less than useful in Radio applications but with a link rate trend > heading for 100’s of Gb and the possibility of FPGA <-> FPGA direct connect > it seems like good future proofing. > > (Brian) In terms of the implementation, IP fragmentation is not supported > at all by the USRP H/W. It added complication and buffering requirements > that were not at all attractive, nor was it something I wanted to advertise > as extensively verified against a vast array of 3rd party networking H/W. > Internally to the RFNoC logic, the implementation was very focused on the > 36kbit native RAM block size in 7 series Xilinx generation and ethernet > jumbo frames. Since we would not fragment CHDR internally to the FPGA then > full assembly/disassembly would have occurred at the periphery of the FPGA > and all logic would then be required to support 2^16 packet sizes had the > implementation fully addressed the max possible architectural CHDR packet > size. > > -Ian > > > On Aug 1, 2018, at 3:38 PM, Leandro Echevarría via USRP-users < > usrp-users@lists.ettus.com> wrote: > > Hello Nick, > > That's the spirit of my question: why couldn't you break up a CHDR packet > over multiple Ethernet frames? I understand it is common for Ethernet to > break up an IP packet (which would also have a 16-bits field for packet > size in its header) into fragments, but limiting the size of the CHDR > packet itself to fit in just one Ethernet frame beforehand seems to be > something thought out well before the data arriving the network interface. > > And Brian: in my case I am indeed using jumbo frames, as my MTU is 9000 > bytes. > > Regards, > > Leo > > > > On Wed, Aug 1, 2018 at 5:27 PM Brian Padalino <bpadal...@gmail.com> wrote: > >> On Wed, Aug 1, 2018 at 3:59 PM Nick Foster via USRP-users < >> usrp-users@lists.ettus.com> wrote: >> >>> That's the MTU of your network interface limiting the CHDR packet size. >>> Can't break up CHDR packets over multiple network packets. >>> >> >> Is the last statement that CHDR can't break over multiple network packets >> a limitation of the implementation or the technology in general? UDP can >> have up to 64k (approximately) bytes for payload size. >> >> Could Super Jumbo frames, if supported by the network, alleviate this >> problem - or is it elsewhere in the system for the limitation as well. >> >> Brian >> > _______________________________________________ > > > USRP-users mailing list > USRP-users@lists.ettus.com > http://lists.ettus.com/mailman/listinfo/usrp-users_lists.ettus.com > >
_______________________________________________ USRP-users mailing list USRP-users@lists.ettus.com http://lists.ettus.com/mailman/listinfo/usrp-users_lists.ettus.com
