network configuration for Starwind MPIO?

iscsi · Sat May 28, 2011 6:16 pm

Hello,
I am thinking about implementing Starwind software first as test product. But i am stuck on how to configure network layer to have enough throughput over 1Gbit ports. I assumed 2 scenarios:
1. Starwind server connected to 1Gbit switch with 3x 1Gbit NICs (all NICs used for iscsi traffic). To the same switch ESXi host with 2x 1Gbit NICs connected.
2. Starwind server connected to 1Gbit switch with 4x 1Gbit NICs (all used for iscsi). To the same switch would be second switch connected (with 2x 1Gbit ports used). To this second port ESXi host with 2x 1Gbit NICs would be connected.

I don't know how should be these network parts (marked with red) configured, so MPIO works as designed. The only thing I would like to achieve is that I have enough throughput if needed by my servers. At this moment I do not plan HA setup (just single box).

What IP's should be used on Starwind ports? all in different subnets? If yes, then which IP's should be configured on ESXi servers to work correctly with MPIO?

Network switches will be cisco or HP (does this make any difference?), both will be managed switches.

Thank you in advance for your help.

Mon May 30, 2011 2:55 pm

Both your interconnection diagrams are wrong. The one with two chained or stacked switches adds extra latency for packet re-route and has multiple points of failure (both switches). Singe switch overheated or just went FUBAR should result whole system going down. Second one with three links between switch and StarWind storage node has zero sense. How do you expect switch routing packets between 2 IPs on ESX and 3 IPs on StarWind? To have properly configured network you should have something similar to either first diagram attached below (the one with one switch) or second (with two switches). First one has MPIO configured with one switch being single point of failure and second one has fully redundant MPIO. Assign two NICs on ESX with IPs for example 192.168.0.1 and 192.168.1.1 and second pair on StarWind side with IPs 192.168.0.2 and 192.168.1.2.

Mon May 30, 2011 2:55 pm

Diagram 1 (MPIO configured with one switch).

Mon May 30, 2011 2:56 pm

Diagram 2 (MPIO configured with a pair of switches).

iscsi · Tue May 31, 2011 6:10 am

Anton,
thank you very much for your answer and your comments.
Indeed in that case (when only one switch is in place) i will connect equal number of paths to both esxi and starwind. That's clear. But second scenario is not clear to me. Second case assumes that esxi servers (3) will be located ~50m from starwind server. So i need to find out how to configure switches and host to provide over 1Gbit throughput if needed. Can you advise something?

Thu Jun 02, 2011 8:01 pm

What you need is called GbE repeater. There are plenty of them on the market. Unfortunately we did not had any chance to try them in the lab and we're not affiliated with any hardware vendor either so I cannot recommend anything in particular. Google is your friend here

If you really need to stick with multi-switch model make sure you have FOUR switches chained to form two fully redundant MPIO links. Something like:

ESX <--> switch 1 <---> switch 2 <---> StarWind
ESX <--> switch 3 <---> switch 4 <---> StarWind

Should work fine. With extra hop and increased latency but fine.

camealy · Mon Jul 11, 2011 8:01 pm

anton (staff) wrote:Diagram 2 (MPIO configured with a pair of switches).

In this MPIO configuration, would you use two different subnets? We have been deploying with two switches in redundant ring stacks with one subnet for iSCSI traffic and then one for the crossover link between Starwind HA members.

Constantin (staff) · Tue Jul 12, 2011 12:32 pm

Offer from Anton to use two switches is rather way to improve your redundancy and avoid single point of failure everywhere, when possible.
Also I wouldn`t use different subnets for iSCSI, but using one subnet for iSCSI and one for sync - is rather good idea.

camealy · Tue Jul 12, 2011 12:36 pm

So do you use the same subnet, different IP's in that subnet, two nics, and two switches? Wouldn't there be path confusion?

Thanks,

Kurt

Constantin (staff) · Tue Jul 12, 2011 2:22 pm

No, all will work OK, even without switches. )

Skyrod7 · Fri Jul 15, 2011 4:13 am

Sorry for silly question, but I still don't understand - Should we use teaming or two separate nics with a different IPs?

With two separate nics Starwind Server will be running as active-active or active-passive storage?
Thanks

Thu Jul 21, 2011 9:15 am

Skyrod7 wrote:Sorry for silly question, but I still don't understand - Should we use teaming or two separate nics with a different IPs?

With two separate nics Starwind Server will be running as active-active or active-passive storage?
Thanks

It's up to you to choose using Teaming or not, it depends on the performance it will provide. You have to test both and decide what to use.
Actice-Active mode or Active-Passive mode do not depend on the type of connection you use (teaming or not) but on the target you use, HA targets are used in Active-Active mode usually, and can be used in Active-Passive mode too. It depends on the MPIO policy you will choose in the Client side (initiator). Mirror Targets can be used only in Active-Passive mode.
The same for Single targets (Basic), if you have two NICs used for connection, in the client side you can choose to have a failover between the NICs once you enable the multipath, or to have them running in Round Robin mode (both active pathes), but this is not a Active-active or active-passive storage this is only the way you use the paths.