3750G performance issues....

rgorman · ‎05-04-2006

We recently have implemented a new switch network based on the 3750G switches. We have 3 separate stacks of switches. 2 stacks of 2x3750G switches are used for our server farms. We also have another stack of 2x3750G switches for a Core. The server stacks are each hooked up to the Core with an 8 line etherchannel (4 lines per switch in the stack for fault tolerance). Servers are hooked up across both switches in the stack using network teaming of the two network cards (HP DL380 servers for the most part). We have our main backup server hooked directly up to the Core stack (with the thought that it was more central to suck the backup data down from the servers). We are not seeing the performance we were expecting out of this setup. We were not responsible for the actual switch configuration, that was subcontracted out to a 3rd party supplier of the equipment. An early copy of the config for the switches seemed to indicate that there was going to be LACP turned on for the etherchannels going from the server stacks to the core switch, but upon final review of the implemented configs there appears to be no LACP turned on (channel-group # mode on). When I review the port statistics with the Cisco Network Assistant it appears as though only single etherchannel lines are being used. I know the etherchannel is setup in a cross-stack format but is it still possible to use LACP to load balance the traffic across all the lines in the etherchannel?

I can provide copies of the running config and possibly a visio diagram of the setup if it would help clarify things...

Any help would be appreciated.

Rob

dominic.caron · ‎05-04-2006

Hi Rob,

I think LACP is only for channel negociation between the switch. If you use etherchannel, a communication between two hosts will always use the same physical link in the group.

EtherChannel balances the traffic load across the links in a channel through the reduction of part of the binary pattern that the addresses in the frame form to a numerical value that selects one of the links in the channel. EtherChannel load balancing can use MAC addresses or IP addresses, source or destination addresses, or both source and destination addresses. The mode applies to all EtherChannels that are configured on the switch. You configure the load balancing and forwarding method with use of the

port-channel load-balance {dst-ip | dst-mac | src-dst-ip | src-dst-mac | src-ip | src-mac}

global configuration command.

The 3750 platform it realy feature rich but I think this behavior is in line with IEEE standars. If you realy need more bandwith between two host, look at 10Gbit ethernet.

Fernando_Meza · ‎05-04-2006

Adding a bit more .. it is very important for you to determine the most utilised hosts on the network .. because depending of the type of load balancing been configured . most traffic could be traveling over the same physical link within the EtherChannel. i.e if you have load balancing dst-mac configured and most of the traffic is directed to a router ... then all traffic will use the same physical link ... and the other links on the ethernel channel will be under-utilised causing a bottle neck and as a result a performance degradation.

In regards to LACP an PAgp these protocols are used for negotiation and automatisation of the config but the load itself is configured manually .. The ethernechannel can be created without the use of these protocols by forcing 'on' on both sides of the link.

I hope it helps ... please rate if it does !!!

rgorman · ‎05-05-2006

I will look into that port-channel load-balance command and see if I can set that. I didn't see it in my current config so it probably isn't set.

The main thing with performance that has been concerning me is that we used to use HP Procurve switches (2848's which are 10/100/1000) and we saw better overall throughput. Our backup jobs would normally finish by 8:00am but now they don't finish until 2:00pm. Nothing else has changed with regard to how the backups are run with the exception of the replacement of the switch gear.

I noticed in the documentation of the 3750's that they use Store-And-Forward by default. Could that possibly be affecting the overall throughput on the network? I am not too sure what the HP's used (they were pretty simple switches so I assumed it was cut-through).

My partner and I are quite baffled by the performance issues we have experienced so far. Any additional suggestions would be much appreciated.

rgorman · ‎05-05-2006

I will look into that port-channel load-balance command and see if I can set that. I didn't see it in my current config so it probably isn't set.

The main thing with performance that has been concerning me is that we used to use HP Procurve switches (2848's which are 10/100/1000) and we saw better overall throughput. Our backup jobs would normally finish by 8:00am but now they don't finish until 2:00pm. Nothing else has changed with regard to how the backups are run with the exception of the replacement of the switch gear.

I noticed in the documentation of the 3750's that they use Store-And-Forward by default. Could that possibly be affecting the overall throughput on the network? I am not too sure what the HP's used (they were pretty simple switches so I assumed it was cut-through).

My partner and I are quite baffled by the performance issues we have experienced so far. Any additional suggestions would be much appreciated.