I'm building a clustered file server using WSFC, vSAN, and iSCSI. Storage Replica causes a major performance drop even though the network is very fast when replication is disabled. I've tested both synchronous and asynchronous replication, and the latency between sites is under 1 ms. Configuring the WSFC replication also hasn't been straightforward. What setup or alternative would you recommend?
2 Answers
DFS Replication may be worth considering if the goal is keeping folders available at multiple locations rather than maintaining a block-level mirror. It works at the file level and is not a replacement for a true active/passive clustered disk, but it can be more practical for shared documents when the application can tolerate replication delay. Be careful with files that are edited concurrently or held open, since conflicts and replicated changes can become an issue. For a highly available file server, validate whether the workload needs continuous failover, namespace-based access, or simply a secondary copy—those requirements lead to different designs.
Synchronous Storage Replica adds replication latency to every write because the destination has to acknowledge the change before the source confirms the I/O to the application. Even with sub-millisecond latency, the extra network, protocol, and storage-processing overhead can be significant, especially with small or write-heavy file-server workloads. Asynchronous mode should avoid blocking every write, but it can still consume bandwidth and storage I/O. I’d benchmark local disk latency, replica throughput, queue depth, and the vSAN/iSCSI path separately before assuming the network is the bottleneck.
I tested both synchronous and asynchronous modes and saw roughly the same slowdown. The sites are less than 1 ms apart, so I’m now checking the storage and replica queues rather than just network throughput.
That’s my concern with DFS Replication: files can change while they’re being replicated, and it doesn’t provide the same block-level consistency as Storage Replica. I’d only use it if the applications and conflict behavior were well understood.