p4lang
Performance discussion.
Hi， I'm a little confused about packets and folding pipes. One pipeline processes 64 elements data at a time. Is this the limitation of Tofino chip? If yes, can the data of 128 elements or 256 elements only be processed through the 2-fold pipeline and the 4-fold pipeline, or can the data of 128 elements and 256 elements not be processed on the switch with only one pipe.
My understanding is that this is indeed due to a chip limitation (authors feel free to correct me). You can only parse a few hundred bytes in one go. For any number of elements, if they fit the parse limit they can be processed in 1 pipe, in 1 go. If they exceed it, you can still have 1 pipe but you have to recirculate and reparse the next N elements, and so on. In switch ML they fold the 3 (unused) pipes to avoid this recirculation which slows down pipe 0, in order to get close to the 100G line rate.
