3-node HA and RAID0

[kevrags]

We have a 3 node license, and are building up the 3rd node to put into our current 2 node cluster. I've read that RAID 0 is recommended due to the higher degree of fault tolerance. It's anyone doing this? I get the rationale, I just have a hard time getting comfortable with it, just thinking about resyncs on our 16TB license for any drive failure or corruption.

Thanks,

Kevin

[anton (staff)]

RAID0 is for performance and not for a fault tolerance. So we build a RAID 0+1 (don't confuse with 1+0 aka 10) with every node having RAID0 and having RAID1 between the nodes (now 3 with a option for 4 with upcoming V8). So with 2 or 3 active nodes left it's easy to re-provision the whole node creating RAID0 from scratch then to re-build the RAID using healthy drives left (on a single node).

[kevrags]

Hi Anton,

Understand the reasoning, just wondering if anyone is actually doing this in a production environment.

Thanks,

Kevin

[anton (staff)]

17% of customers running StarWind bare metal and 31% of Native SAN for Hyper-V / vSAN run it in a 3-node config.

[jtmroczek]

Kevin:

I have 2-node HA servicing Hyper-V workloads. We are using RAID 0 without issues.

No matter how you slice it a full resync of 16TB is going to take some time. If you have a good 10GbE network, preferably with multiple links, it should be tolerable. We have dual 10GbE links for the sync channel. 4TB (our largest LUN) resyncs take about 45 minutes.

As always, test test test.

~joe

We have a 3 node license, and are building up the 3rd node to put into our current 2 node cluster. I've read that RAID 0 is recommended due to the higher degree of fault tolerance. It's anyone doing this? I get the rationale, I just have a hard time getting comfortable with it, just thinking about resyncs on our 16TB license for any drive failure or corruption.

Thanks,

Kevin

[Anatoly (staff)]

jtmroczek, thank you very much for the feedback - we really appreciate it.

kevrags, we have pretty a lot of customers that are more then happy with running Sw on RAID 0 config, so feel free to use such config in your environment as well.

[anton (staff)]

...and we should assume HA is for a Business Continuity and async replication is for Disaster Recovery. So VMs or LUNs should be replicated or VMs backed up.

[kevrags]

jtmroczek,

Thanks for the post. We currently only have a single 10Gbe sync channel, mainly because the R710s we use only have 2 x8 slots, and the RAID controller houses the other one. Although we will probably add them later, as the x4 slots use the x8 form factor.

The other issue is how to network 3 nodes with 10Gbe and not break the bank, as our 2 node systems are directly connected. Probably the new Netgear 10Gbe, but we will have to switch to copper from our current SFP+ adapters. ~$1,800 for 2 8 port switches.

Kevin

[anton (staff)]

For 3-way design you can create a pseudo-Token Ring. Throw in a pair of a 10 GbE cards into every server (no much need in hardware RAID controller
as Windows can handle RAID0 in software just fine) and connect the servers in a "circle". So

... C_Port2 <- A_Port1, A_Port2 <-> B_Port1, B_Port2 <-> C_Port1, C_Port2 -> A_Port1 ...

That design would not require a switch as long as switch is more expensive then a 3 10 GbE cards. Also proposes design is FASTER then using a single 10 GbE port on a box.

jtmroczek,

Thanks for the post. We currently only have a single 10Gbe sync channel, mainly because the R710s we use only have 2 x8 slots, and the RAID controller houses the other one. Although we will probably add them later, as the x4 slots use the x8 form factor.

The other issue is how to network 3 nodes with 10Gbe and not break the bank, as our 2 node systems are directly connected. Probably the new Netgear 10Gbe, but we will have to switch to copper from our current SFP+ adapters. ~$1,800 for 2 8 port switches.

Kevin

[chrisjacob]

I have read before that striping data only on the first server of a cluster can lead to corrupt the data that would be copied over to the second server in the cluster. HA should be accounted for on all servers, not just the cluster. I would look at doing the OBR10 on both servers to ensure complete data integrity.




Thanks

[anton (staff)]

OBR10 is not going to help here. If app had managed to corrupt content that content would be delegated to all partners in a LU mirror group. RAID level below synchronous replication is irrelevant... Backup would help. HA is for BC and asynchronous replication is for DR.

[BillArchway]

Kevin:

I have 2-node HA servicing Hyper-V workloads. We are using RAID 0 without issues.

No matter how you slice it a full resync of 16TB is going to take some time. If you have a good 10GbE network, preferably with multiple links, it should be tolerable. We have dual 10GbE links for the sync channel. 4TB (our largest LUN) resyncs take about 45 minutes.

As always, test test test.

~joe

Wow. I'm not getting anywhere close to that sync performance. We have a 6tb LUN that just took 10 hours to sync. We're running raid-10 with a single direct connected pair of 10g nics. I see sync nic utilization at about 25%, or apx 2.5gb.

[anton (staff)]

Then I'd suggest to schedule a remote session with our staff and see what's wrong with your config, what could be improved / fixed etc.

Kevin:

I have 2-node HA servicing Hyper-V workloads. We are using RAID 0 without issues.

No matter how you slice it a full resync of 16TB is going to take some time. If you have a good 10GbE network, preferably with multiple links, it should be tolerable. We have dual 10GbE links for the sync channel. 4TB (our largest LUN) resyncs take about 45 minutes.

As always, test test test.

~joe

Wow. I'm not getting anywhere close to that sync performance. We have a 6tb LUN that just took 10 hours to sync. We're running raid-10 with a single direct connected pair of 10g nics. I see sync nic utilization at about 25%, or apx 2.5gb.