Hello, Miguel Well, hopefully Simon-Bernard's suggestion helps.
If not, you could also try to do this (on basically any client OS) with L3 routing, I think. For that you set up the backup server's "public IP address" (one published in DNS) on its loopback interface alias. The NIC interfaces have some other IP's (and MAC addresses). These IP's are published to clients as host-routes (/32) with the same metric as best routes to the backup server's "public IP address". This way the client might begin alternating the "gateway" it uses to deliver packets to the backup server's public IP address. In L2 that is, the client sends Ethernet frames to different MACs and thus the switch load-balances them. You can use any means, from static routing rules on the client (probably during testing or in small deployments) to RIPv2, OSPF and other local routing announcements. In my Solaris experience, at least default routes are often alternating like this. If host routes don't, you might set up some script to change these routes' metrics back and forth :) Some other load-balancing implications include this: if you use different IP subnets for the client's and server's "gateway" NICs, then the client packets destined to a certain gateway would originate on a client's certain interface. This may happen to be useful in setups where there is no etherchannel (i.e. when two switches used for redundancy). All-in-all, I won't estimate whether you can exceed 1Gbit this way, but you can try anyway :^) //Jim [email protected] пишет: > Hi Jim > > Thanks a lot for your valuable feedback. > Let me try to explain what the problem is and how I would like to solve it. > > Consider a backup server A and backup client B, interconnected via a gigabit > switch. > In order to achieve a throughput higher than 1 gigabit, both server A and > client B have an EtherChannel configured. > However, there is a problem that prevents me to achieve a bandwidth higher > than 1 gigabit. > The physical switch is a layer 3 switch, which is only capable of doing > load-balancing based on either layer 3 info (IP) or layer 2 info (Ethernet). > In order to break the 1 gigabit barrier, I would have a give the backup > server multiple IP addresses (of which each one is guaranteed to be > redirected to a different physical port on the switch). > > The problem here is that whenever the backup client sets up multiple TCP > connections to the server, it only uses 1 IP address of the server (It > performs one DNS query and uses the result for each of the connections to set > up). As a result all TCP sessions will have the same destination address, > resulting in a maximum bandwidth of 1 gigabit since all traffic will hitting > only one physical switch port on the destination side. > I was hoping to solve this with ipf locally on the client so that each TCP > session could get directed to a different destination IP address. -- +============================================================+ | | | Климов Евгений, Jim Klimov | | технический директор CTO | | ЗАО "ЦОС и ВТ" JSC COS&HT | | | | +7-903-7705859 (cellular) mailto:[email protected] | | CC:[email protected],[email protected] | +============================================================+ | () ascii ribbon campaign - against html mail | | /\ - against microsoft attachments | +============================================================+
