New build/testing

[WolfR1der]

We have two Win2016 servers and after exhausting research as to how to set up clustered volumes for Hyper-V with MS Clustering we found that not only can you not do it without a third server but Starwinds will do what we're looking for.

Note that we are sticking with the free versions of Starwinds
On initial set up I've created a 2.5TB iSCSI volume and replicated it to the mirror server. This replication process is taking a very long time. Network activity shows an average of 907Mbps passing through a NIC team of two 1Gbps NICs. Would putting a 10Gb NIC in have a noticeable impact? The HDD subsystem (RAID6) holds around 6% max utilization so that's nowhere near a bottleneck. Since this setup will be for Remote Desktop Services I'd like to make sure synchronization is never a hindrance to operation.

Also would it be better to split the HA disk into two images to avoid split brain issues or does this software mitigate the problem of two server accessing the same volume issues?

[Boris (staff)]

It is possible to setup MS Failover Cluster with two nodes, no third node is required.

On initial set up I've created a 2.5TB iSCSI volume and replicated it to the mirror server. This replication process is taking a very long time.

To avoid long synchronization time on disk creation, create your disk as a small one (a couple of GB will be fine), replicate it to the partner node and then extend it to the required size. This wasy it will be much faster.

Network activity shows an average of 907Mbps passing through a NIC team of two 1Gbps NICs.

Never use NIC teaming or any kind of NIC aggregation for StarWind's iSCSI or Sync connections. Better use 2 separate Sync interfaces to avoid additional latency on link aggregation. Moreover, your NIC team might be configured with one link active and the other one passive, which results in having only half of the potential throughput.

Would putting a 10Gb NIC in have a noticeable impact?

It will.

The HDD subsystem (RAID6) holds around 6% max utilization so that's nowhere near a bottleneck.

You would never expect decent performance for write operations on RAID6. Check https://knowledgebase.starwindsoftware. ... ssd-disks/ for recommended settings (RAID10 on HDD).

Also would it be better to split the HA disk into two images to avoid split brain issues or does this software mitigate the problem of two server accessing the same volume issues?

It is better to split the workload between the nodes, so creating at least two smaller cluster shared volumes instead of one large volume will be better, as you will be able to distribute the CSV ownership between the nodes, and the same will apply to VMs running inside the cluster. It is recommended that VMs running off a CSV should be owned by the same cluster node as the CSV they are running off. This has nothing to do with avoiding split brain, as for avoiding split brain it is important to add heartbeat links on the iSCSI and the management interfaces.

[WolfR1der]

It is possible to setup MS Failover Cluster with two nodes, no third node is required.

With your software correct. With Microsoft's Clustering service there is a need otherwise your software wouldn't be necessary.

To avoid long synchronization time on disk creation, create your disk as a small one (a couple of GB will be fine), replicate it to the partner node and then extend it to the required size. This wasy it will be much faster.

Thanks. Is there a way to do this via Powershell?

Never use NIC teaming or any kind of NIC aggregation for StarWind's iSCSI or Sync connections. Better use 2 separate Sync interfaces to avoid additional latency on link aggregation. Moreover, your NIC team might be configured with one link active and the other one passive, which results in having only half of the potential throughput.

The current NIC team is using LACP with dynamic load balancing. Max traffic I saw during the synch was averaging 907Mbps

You would never expect decent performance for write operations on RAID6. Check https://knowledgebase.starwindsoftware. ... ssd-disks/ for recommended settings (RAID10 on HDD).

What I was alluding to was that the RAID system was nowhere near taxed because the network system was a bottleneck.

It is better to split the workload between the nodes, so creating at least two smaller cluster shared volumes instead of one large volume will be better, as you will be able to distribute the CSV ownership between the nodes, and the same will apply to VMs running inside the cluster. It is recommended that VMs running off a CSV should be owned by the same cluster node as the CSV they are running off. This has nothing to do with avoiding split brain, as for avoiding split brain it is important to add heartbeat links on the iSCSI and the management interfaces.

I will try that.

[Boris (staff)]

Thanks. Is there a way to do this via Powershell?

Sure. Initially, in your device creation script you need to indicate its size as some 1024MB, for instance, and after that you would include some adjusted part of the ExtendDevice.ps1 script from the StarWindX PowerShell samples folder.

The current NIC team is using LACP with dynamic load balancing

No form of NIC teaming is good for iSCSI connections. Although technically it is possible to use NIC teaming, I would not recommend doing so.