There have been quite a number of search expressions here on by blog, so decided to answer them. Not that many are clear but I am trying to do my best to “decode” them.
Witness related questions
Is witness required at all? When is it needed?
The witness component is required for the automatic switchover (ASO) to function. The term “peer persistence” a generic term which expresses sync replication with ASO together.
Wintess is not required in sync relationship, so if ASO is not required, no witness is needed. In cases like this the loss of the active volume owner will not bring online the volume on the surviving array, it must be done manually as no ASO will be available.
How many witness components are required if I will setup two pairs of PP?
Two. Witness cannot be shared between arrays in different PP relationship. Furthermore needless to say, but in one PP pair there can be one and only one witness.
HA related questions
There are two distinct scenarios.
HA event within one array
A type of event which can happen if an unrecoverable problem is impacting the active controller in an array. Cause can be anything from malfunctioning NVDIMM module up to dead capacitor, anything.
Failover from active to stanby controller must be happening below 30 seconds and based on me measurements – NOS 5.0.7 – it is done under 20 seconds. This is something HPE is trying to shorthen with each new release. Of course during the failover process no write will be lost since the host with the write request will not receive an ACK as long as the failover is in progress and a new active controller is not selected. This is a major event since in normal operation of a Nimble array it is coded into its DNA, to only ack a write, if that has been acknowledged by the standby controller too. So once the standby controller stores that write in its NVDIMM module, will reply back to active controller and that active controller to the host with success.
In a case when there is only one working controller and no standby one this is not happening and any issue which affects the only controller might result in data loss. This is VERY unlikely, since the capacitor in it will still power the NVDIMM module during the copy process of the DRAM chip content to the flash chips on the module other side. Anyway this is a situation which must be avoided at all cost, so make sure you have support!
HA event between two arrays
If two arrays are in one PP relationship, the incoming write actually stored 4 times. All happens pretty much the same as in single array design, but ACK is sent back to the host only if the active array, active controller and the downstream array, active controller stores the write in NVDIMM. Furthermore these active controllers – one in each array – will also ensure their passive controller pair to have the request as well. But notice the major difference here:
- in non PP relationship ACK comes when active and passive controller in one array have the write stored in NVDIMM.
- in a PP relationship ACK comes when active controller in upstream array and the active controller in the downstream array have the write stored in NVDIMM. – for a give volume at least.
HPE Nimble Peer Persistence whitepaper
There is not a document so far like that. Anyway it is pretty much clear at least around requirements. 5ms RTT between arrays, 250ms between arrays and the witness. IP transport only.
iSCSI replication configuration
Replication is not iSCSI, it is proprietary.
Direct attach iSCSI
Not supported. Direct connect – so to connect a host directly to the array’s port – is supported only if FC is used.
Peer persistence limitations
Maximum of 128 replicated volumes and these volumes must be presented to the host(s) using the same protocol. So if a host on site A uses FC to reach a volume, the same volume on Site B must be exported over FC as well. There is no mix n match here.
Different models in peer persisence
Lower model cannot sync replicate to a higher category model – for now. So AF20 will not sync replicate to AF40.
Higher category model cannot sync replicate to lower category model.
Two models in the same category, but different subcategory can sync replicate. So an HF20C can work with an HF20.