On Wed, 2013-07-10 at 15:44 -0700, Dave Taht wrote: > I'd made a suggestion elsewhere that TSQ migrate down in size from 128k to > lower as the number of active flows increased. Something like > tcp_limit_output_size = max((2*BQL's limit)/(number of flows),mtu) > > but I realize now that tcp has no idea what interface it's going out > at any given > time... still I'm on a quest to minimize latency and let offloads still work..
At Google we tried to plug something at the time TX completion happens (tcp_wfree()) The more time skb are waiting on qdisc, the less tcp_limit_output_size should be for the TCP flow. But tcp_limit_output_size had to be per socket tunable instead of global. Experiments showed no real improvement over existing TCP behavior. The tcp_tso_should_defer() was kind of fixed lately [1] anyway, and we plan to upstream another patch in this function to better preserve ACK clocking. [1] : http://git.kernel.org/cgit/linux/kernel/git/davem/net-next.git/commit/?id=f4541d60a449afd40448b06496dcd510f505928e _______________________________________________ Codel mailing list [email protected] https://lists.bufferbloat.net/listinfo/codel
