I used to be an expert of TCP (well, in college), but I've forgotten lots things about it, LoL.... But I am sure that the SACK TCP can handle the out of order issue. It will work more "smoothly" if the packets arrive in order though. But the overall throughput of "SACK TCP with LACP" is unknown for me. Maybe it's an interesting issue for researchers. I will survey the related documents about it. And thank you, you inspire me a lot!!
2012/4/7 Ethan Jackson <[email protected]> > Please keep replies on the list. > > > That's a good point for why a TCP flow cannot be distributed > > to different links simultaneously. Thanks a lot. > > But the TCP SACK mode can handle the out of order issue, isn't it? > > I'm not a tcp expert. My understanding is that out of order packets > typically cause tcp backoffs. I may be wrong on this point, but in > general networking equipment tries to deliver packets in order for > this reason. Even if tcp can handle it gracefully, a bonding > implementation that consistently delivers packets out of order is > incorrect. > > Ethan > > > > > > 2012/4/7 Ethan Jackson <[email protected]> > >> > >> >> OVS can't distribute packets of the same flow to different > interfaces. > >> > >> As far as I know, this is typical of all bonding implementations. If > >> you send packets from the same flow down different links, they may > >> arrive at the destination out of order. This can wreak havoc on tcp > >> stacks. > >> > >> >> OVS can't really achieve balance/fairness between flows. > >> >> OVS can't dynamically switch a flow between different interfaces on > the > >> >> fly based on the bandwidth utilization. > >> > >> This isn't true, if one slave is receiving quite a bit more traffic > >> than another, some flows will be rebalanced to the less loaded one. > >> By default this is done once every 10 seconds, but that's > >> configurable. > >> > >> >> So it means the LACP implemented in OVS is not mature? > >> >> Or it's just the nature of LACP? It needs extra QoS scheme to achieve > >> >> real balance and full bandwidth utilization? > >> > >> This is just the nature of bonding implementations. > >> > >> Ethan > >> _______________________________________________ > >> discuss mailing list > >> [email protected] > >> http://openvswitch.org/mailman/listinfo/discuss > > > > >
_______________________________________________ discuss mailing list [email protected] http://openvswitch.org/mailman/listinfo/discuss
