Switch Assisted Load Balancing Revisited

[Jason Scott]

I'm posting again in the hope that some more people will reply, as I'm sure this has to be working for others.

I want to setup load balancing for both incoming and outgoing traffic from an HP 740 server with two HP Nics connected to an HP 5308 switch.

At present it appears that only 100MB of this 200MB trunk is being used. I have tried combinations of the following:

TRUNK, FEC, LACP Static and LACP Dynamic.

The last option was a long shot because the server teaming software documentation mentioned LACP must be in static mode.

Could someone also clarify the options on HP teaming software for MAC address or IP Address load balancing. I understand MAC address balancing when a routed gateway exists between server and clients, will result in one sided load balancing. However if the teaming software is set to IP address balancing, how does the layer 2 switch handle that?

Can anyone confirm that they have an HP switch load balancing above 100MB to a server?

[Steven Koutstaal]

hi Jason (sorry for the later reply, I only just saw this forum and thought it was a nice opportunity to learn something about HP's switches and help others).

Basically, the procurves do _destination_ based load balancing. This is either MAC- or IP-address based.

As long as the server has only one IP-(switch trunk) or MAC address(server trunk) this will come down to the fact that it will always choose the same physical link to go to the server, thus restricting you to 100M incoming.

I found this bit of info on the HP website;
http://www.hp.com/rnd/library/pdf/59692372.pdf

cheers,
/steven

PS: Extreme Networks can do loadbalancing on MAC/IPaddr/TCPorUDP Port for src and dst, giving you a good spread..

[Ardon]

Steven,

Sorry but I have to correct you on your "Basically, the procurves do _destination_ based load balancing. This is either MAC- or IP-address based.

As long as the server has only one IP-(switch trunk) or MAC address(server trunk) this will come down to the fact that it will

always choose the same physical link to go to the server, thus restricting you to 100M incoming." statement.

If you use FEC/LACP/HP Trunking all are Layer 2 based SA/DA meaning that the ProCurve Switch will Load Balance taken into consideration BOTH Source AND Destination MAC Address (Logical X-Or operation). Mind you that Trunking (unlike Meshing) is NOT dynamic. With Trunking a certain combination of SA/DA will ALWAYS take the same physical link no matter what load already on that link.

Regards, Ardon

[Richard Pearl, Jr.]

I have a similar question. If I want to use SLB teaming on a Compaq ProLiant server, should I set my HP ProCurve 5308xl to use FEC for the ports connected to the "teamed" NICs?

[Jason Scott]

Rick, it depends on the Proliant setup on the software drivers that control the teaming. What we have found now with our servers is that we do not need to configure any trunking on the switch as the server handles things outbound. Inbound it doesn't seem to matter because the switches seem incapable of doing anything intelligent with that traffic.

[Richard Pearl, Jr.]

Jason,
Thanks for the reply.

If the switch isn't configured for trunking and you use any form of teaming on the server (my preference is SLB for throughput reasons), would that not cause a switch loop?

The way I understand teaming is that all the NICs configured in the "team" would then take on the IP adddress and MAC address of the primary NIC in the team. Am I understanding that incorrectly?

One of the main reasons we are looking at teaming is to provide a bigger "pipe" for inbound traffic to the server(s), but if teaming won't do that, then we will need to rethink how to proceed.

Any suggestions on how to accomplish this?

[OLARU Dan]

We tried hard with this teaming stuff between our backup server and an HP4108GL switch. Didn't work. Gave up.

Jason is right: "Inbound it doesn't seem to matter because the switches seem incapable of doing anything intelligent with that traffic"

So we purchased an HP2848 L3 switch who has 48 x 10/100/1000 ports, plus 4 nice dual personality ports. We ordered certified CAT6 patch cords to make sure we can go on 1Gbps and we'll try to put our 1Gbps servers at work.

I have one question, tho: isn't the bandwidth of the internal bus of the servers going to be the new bottleneck?