Proliant Network Teaming - Windows Load Balancing - Procurve Trunking

[Darren Etheridge_2]

I have a few Proliant 380 SQL servers that I would like to get the best network thruput possible.
I have teamed the onboard GB nics, now I guess my question is:
1. Do I also need to enable Windows Load Balancing?

2. Do I need to create a trunk for these connections in my procurve Switch.


I've read quite a few questions & answers regarding this, but no clear cut yes or no answers :)

Thank you for your time!

D

[Lmm_1 · Accepted solution]

Darren, it depends of what your goal is. If you are looking for bandwidht, I would recommend using 802.3ad Teaming mode, this will require an LACP channel to be configured on the procurve switch:

trunk x,y trk1 lacp

.Lmm

[rick jones]

Best throughput possible in what sense? Single connection throughput? Multiple connections between two machines? Multiple connections from one machine to many other machines?

[Darren Etheridge_2]

Multiple Connections from one machine to another machine thrua procurve switch

[Darren Etheridge_2]

It seems the Procurve switch is automatically setting up a trunk (Dyn1) for the teamed connections... is this prefered ?

[Andrew Pearce]

Personally, I prefer to manually create the trunks using the command in a previuos thread.
trunk 1,2 trk1 lacp

That assumes that the ports 1 and 2 are connected to the server. and you dont have a static trk1 already configured.

Also if you have more than one procurve switch you can use dt-lacp.But lets not run before you can walk.

[rick jones]

Multiple connections between the same two machines suggests a few things:

1) load balancing algorithms in both the host(s) and the switch which include layer four protocol addressing information (ie TCP port numbers)

2) using 10G Ethernet rather than 1G and eschew the whole trunking thing save for HA.

If either the host or the switch cannot include layer four addressing information in the load balancing algorithm you might need to fall-back on:

1a) using multiple IP addresses on the hosts and making certain your multiple connections are spread across those IP addresses

Backing up a bit, and since my forte is more networking than database - it is my understanding that SQL is rather "request response" and not generally bulk throughput. Do you have any stats about the average packet size in/out of your systems serving SQL? Small packet performance may care rather more about stack path length and intra-host load balancing than the bitrate of the link(s) connected to the host.