On 25 May 2015 at 12:09, Kury Nicolas <[email protected]> wrote:
> Hi > > > I have a question about the documentation of OpenDataPlane. > > > http://www.opendataplane.org/wp-content/uploads/2014/01/ODPIntroductionandOverview-2014Jan29.pdf > > > Page 15, there is written : > *Another application and SoC may have to sustain 10 Gbps or 15 Mpps with > just a few **Watt power budget (e.g. max four 1.5 GHz CPU cores), which > would then result to total **cycle budget of ~400 CPU cycles per packet.* > > 400 CPU cycles is quite high, isn't it ? > A modern Xeon CPU can do layer-2 switching (e.g. using OVS) in approx 200 cycles. Most other CPU's don't have to IPC of a Xeon so may need more cycles for this functions (performance also depends on other things, e.g. how actual network I/O is implemented). But layer-2 switching is one of the simplest networking functions. Other functions may perform parsing and classification, a couple of table look-ups, stateful modifications to the packet headers and flow state (e.g. update statistics counters), encapsulation/decapsulation etc. all within the same cycle budget. > When I read description about NetMap, there is written that 10Gbpps (or > 14.88 Mpps) can be reached with a single core running at 900 Mhz (~60 CPU > cycles). > I think this measure excludes any actual processing on the packets. Add in packet parsing and a couple of table look-ups in some large sparse tables and you are at 200 cycles. > Chapter "More details" http://info.iet.unipi.it/~luigi/netmap/ > > Is there anything I didn't understand ? > Yes. The purpose of networking box is not to simply pass through packets from ingress to egress interfaces without any processing. > > Thank you! > Nicolas > > > > > > _______________________________________________ > lng-odp mailing list > [email protected] > https://lists.linaro.org/mailman/listinfo/lng-odp > >
_______________________________________________ lng-odp mailing list [email protected] https://lists.linaro.org/mailman/listinfo/lng-odp
