|Main Archive Page > Month Archives > drbd-user archives|
Matteo Tescione, on 03/06/2010 09:43 PM wrote:
> Hi folks,
> after a little bit of experimenting I SUCCESFULLY created an drbd active/active cluster with scst/iscsi exported to round-robin vsphere initiator.
> configuration is 2 virtual machines with centos5-64 linux-2.6.33 patched and drbd8.3.7. Initiators are vSphere Software iscsi initiator.
> Relevant config in drbd is net section, allow-two-primaries.
> Relevant config in scst.conf is:
> DEVICE rmnet-devel,/dev/drbd0,NV_CACHE,512
> [ASSIGNMENT Default_iqn.2010-02.com.scst:RMnet-devel]
> #DEVICE <device name>,<lun
> DEVICE rmnet-devel,0
> Note that I'm using the same scst.conf and iscsi-scstd.conf in both targets.
> The vSphere initiators are seeing 2 paths for 1 device, switched to Round-robin.
> Since the targets are running virtual with no real hardware, i have no idea at the moment of what kind of performance increase it could bring.
> Vlad, what do you think about? is fileio the best choice here?
Since DRBD sees write requests only when they flushed from the system
cache, write back caching might create a gap for data corruption, so I
wouldn't recommend for your case to use it. Unfortunately, when used on
replicated targets DRBD hooks on not the best level. It should hook
directly with the target to see its writes on the moment they written to
the cache. Only in this case write back caching would be safe with DRBD.
Hopefully, somewhen in the future we will do that. SCST 2.1 is going to
have some features which will greatly simplify this task.
You should add WRITE_THROUGH and delete NV_CACHE options, then you
configuration will be safe. But you then should loose some write
performance. Depending from your storage configuration this loss can be
drbd-user mailing list