Database Reference
In-Depth Information
only when the virtual machine is not using SCSI bus sharing. It is the SCSI bus
sharing required in a SQL Server AlwaysOn Failover Cluster Instance that
prevents the vMotion operations from being supported currently, not the fact the
VM is configured to use RDMs. See VMware KB 1005241.
Although there are no noticeable performance differences between a single VMDK on a
VMFS data store and an RDM, there are important performance considerations and
constraints with using RDMs that need to be considered, such as:
An RDM maps a single LUN to a virtual machine, so each VM will likely
consume multiple LUNs and there will be more LUNs to manage.
More LUNs are required, which may constrain the number of VMs possible as the
maximum number of LUNs per host is currently 256.
It is not possible to perform storage IO quality of service on a pRDM; therefore, a
VM configured to use a pRDM could negatively impact the performance of other
VMs using the same underlying shared storage array.
Can't leverage vSphere features such as Storage vMotion, so it can be more
difficult to balance capacity and performance when using pRDMs and more
difficult to resolve any storage hot spots.
Due to the management overheads, constraints, and VMware feature limitations of using
RDMs, we recommend their use only when absolutely necessary, such as to deploy SQL
FCI; in all other cases, VMDKs should be used. Using VMDKs future proofs your
environment and allows it to benefit from any further advancements that VMware
releases that pertain to VMDKs.
The IO Blender Effect
When you virtualize SQL and consolidate many SQL VMs onto fewer hosts, the amount
of IO per host increases. In addition to the increase in IO per host, in most cases the IO
patterns will also change. Unless you have completely dedicated storage for each SQL
VM, which is not optimal from a cost or performance perspective in most cases, all IO
will at some point become random.
Any time you share storage and have multiple VMs and different IO workloads, the
combined IO pattern is random. Random IO, especially when write biased, can have a
significant impact on storage performance, particularly when RAID (Redundant Array
of Inexpensive or Independent Disks) is used. Grouping similar workloads together can
help improve the blended IO pattern and reduce the burden on storage.
Figure 6.15
shows the impact of combining different IO workload patterns.
