4 ms·
To get a grip around the numbers, i quote from George Neville-Neil's Talk at BSDCan '15 "Measure Twice, Code Once" [1]: - 10 Gbps is 14.8 million 64 byte packe
by sydney6 5y ago
To get a grip around the numbers, i quote from George Neville-Neil's Talk at BSDCan '15 "Measure Twice, Code Once" [1]:
- 10 Gbps is 14.8 million 64 byte packets per second
- 67.5 ns per packet or 200 cycles at 3 GHz
- Cache miss is 32 ns
[1] https://www.youtube.com/watch?v=LE4wMsP7zeA https://www.youtube.com/watch?v=LE4wMsP7zeA
- eqvinox 5y agoYou're not running a continuous stream of 64-byte packets in a home or SME setup. Also, assuming a 1:1 mapping to packet processing is a false dichotomy these days, NICs are doing an unbelievable amount of preprocessing, particularly grouping related packets together.
- sydney6 5y agoNo, of course not. A good starting point for real world performance benchmarking could be e.g. IMIX [1]. The example above represents the solely theroretical worst case as a means to establish a baseline for performance benchmarking. Anyway, if you are referring to HW offloading capabilities of "modern" NIC's, using techniques like LRO would break the "end-to-end"-principle of a router. [1] https://en.wikipedia.org/wiki/Internet_Mix https://en.wikipedia.org/wiki/Internet_Mix