MinRAM : 3236
MaxRAM : 3236
TotalDisk : 260096
While the per disk QoS in Windows Server 2012 R2 is useful in certain scenarios, it
fails to meet the requirements in many scenarios. I cannot actually reserve
performance for storage nor can I centrally manage the policies for shared storage.
Windows Server 2016 changes this with a centralized QoS implementation that is a
feature of the cluster and enables actual reservation of performance. The storage must
be CSV. If you are not using CSV, only the old 2012 R2 style QoS can be used.
However, pretty much every Hyper-V environment uses CSV for its storage today.
Windows Server 2012 R2 implemented maximums for IOPS, which is relatively
simple; however, there was no way to achieve minimum guarantees of performance as
this is a far more difficult problem to solve. Windows Server 2016 tackles this in
various ways; at its core is a new maximum-minimum fair-sharing technology that
ensures that all targets with a minimum policy applied receive that level of
performance, and then additional IOPS are distributed fairly until any maximum that
may also be defined. By using a centralized policy controller, there is no need for
complicated distribution solutions of the actual policies that will be enforced by the
nodes in the cluster.
A key term for QoS is normalized IOPS, which by default is an 8KB block. Anything
less than 8KB is a single, 1 normalized IOPS, while larger I/Os would be multiple
normalized IOPS. For example, a 9KB I/O would be 2 normalized IOPS, a 24KB I/O
would be 3 normalized IOPS, and so on. Note that this is a unit of measurement used
by the various accounting methods that make up storage QoS. It is possible to
reconfigure the normalized IOPS size if 8KB is not suitable for the type of workloads
in your environment. Most organizations, however, should leave the default 8KB.
Three primary components make up the new storage QoS:
1 . On each Hyper-V, compute node is a profiler and rate limiter that tracks utilization
of storage and enforces the maximum IOPS values in the same way that it works in
Windows Server 2012 R2.
2 . An I/O Scheduler is distributed on the storage nodes backing the VHDs (that is,
the Scale-Out File Server). This enforces the minimum IOPS allocations. Note that
in a hyperconverged deployment where the storage and compute are the same
nodes, the I/O Scheduler runs on those hyperconverged nodes in the same way.
This also would apply when the storage is a LUN on a SAN or Storage Spaces
Direct.
3 . The centralized policy manager is a resource in the cluster, which means that it can
move between nodes as required, and is the aggregate point for collecting and
consolidating metrics about storage and driving the management of the policies.
This can be seen as a cluster core resource from within Failover Cluster Manager
or through PowerShell.