ESXi 4.1u1, MPIO & Starwind slow startup/storage rescan

[Constantin (staff)]

Mark, doesn`t matter if you are using 1 subnet or 4. RoundRobin in ESX will anyway use only one path at a time. ESX switches to other paths every 1000IOPS (by default).
But behavior of Windows seems to be be pretty strange to me. As Anton said, we`ll take a look on it, and will provide you with an update.

[mkaishar]

Mark, doesn`t matter if you are using 1 subnet or 4. RoundRobin in ESX will anyway use only one path at a time. ESX switches to other paths every 1000IOPS (by default).
But behavior of Windows seems to be be pretty strange to me. As Anton said, we`ll take a look on it, and will provide you with an update.

Yup...this is why I setup this in rc.local to deal with high iops

esxcli nmp device list | grep ^eui |
while read device ; do
esxcli nmp device setpolicy --psp VMW_PSP_RR --device ${device}
esxcli nmp roundrobin setconfig --type "iops" --iops 3 --device ${device}
done


vmware's response with the use of multiple subnets
http://kb.vmware.com/selfservice/micros ... Id=1024476


I've tested mpio rr with many different iscsi storage vendors: equallogic, emc ax/15xx series, openfiler (open-source), open-e, windows storage server 2003/2008 and starwind...when connecting from ms iscsi or linux iscsi initiators you need mulitple subnets for mpio rr to work and load balance over multiple nics, if you don't do that you will notice all traffic going over primary nic.

emc ax/15xxx iscsi requires their powerpath client to get mpio to work and you still need multiple subnets

the exception is equallogic, with their HIT client mpio works on a single subnet

hope that provides you more details in your testing...thanks.

[Constantin (staff)]

Well, I`m not sure what other vendors recommend, but using separate subnets allow you just split traffic.
About running RR in VMs: of course, it always uses one physical NIC on ESX. The reason it is that we are using RoundRobin LB policy for Datastore, while we don`t use any LB for the NICs. The reason of it is simple - LB policies in ESX are pretty dump, specially for storage. The only really useful policy for storage is "Based on physical NIC load" available in dvSwitch only.
Second option, but I`m not sure if it will work OK for you, configure LACP on switch, SAN and in vSwitch set to use "Based on IP" policy. But as I wrote, I`m not sure if it will work in such config. I would rather use 2 vSwitches.