On 10/08/2015 02:52 PM, Yuchung Cheng wrote:
I would like to add that GRO also cause stretched ACKs (acking up to 64KB), and GRO is crucial to reduce cycles for +10Gbps transfers on some architectures.
And they go back even further than GRO (or LRO) appearing in Linux. They go all the way back to the mid/late 1990s when HP-UX 11 and Solaris 2 started shipping their cousin (Mentat origin) stacks with ACK avoidance heuristics in them. Perhaps even before, but that is the earliest I know of.
GRO, TSO, CKO etc have come into being from the fields made fertile by the combination of ever increasing link speeds, never increasing de jure Ethernet MTUs, no longer increasing processor core frequency, and paying customer demand for performance.
happy benchmarking, rick jones _______________________________________________ aqm mailing list [email protected] https://www.ietf.org/mailman/listinfo/aqm
