A Contributed Article by Lawrence Gavin, Head Geek, SolarWinds, Virtualization & Storage Management
One of the notable financial burdens
of operating a Storage Area Network (SAN) is the expense of Host Bus Adapters
(HBA) needed to allow compute nodes to communicate with the SAN and access
storage resources. A typical cost for an HBA is around $1500, and most systems
need two of them for redundancy.
However, with the advent of Windows Server 2012 (WS2012) and
the new Server Message Block (SMB) v3 file services that are available, you can
now make the storage in your SAN available to compute resources on the network
without the expense of installing HBAs in every system-by placing a WS2012 file
server cluster in front of the SAN.
In addition, by eliminating HBAs from every compute node, you
will also significantly reduce the number of switch ports that would need to
exist in the fabric-potentially eliminating the need for additional acquisition
of switches altogether.
Better Data Throughput for Less
At the basic level, WS2012 servers are each equipped with a
pair of 16GBit/sec Fibre Channel HBAs, just like any other compute node. For more
throughput, you can install additional HBAs in the file services clusters and increase
the throughput from the 32GBit/sec available with just a pair of HBAs, to
64GBit/sec by installing four HBAs.
However, with only the WS2012 file services cluster nodes connecting
directly to the SAN, you also have many additional options for increasing
throughput from the slower Fibre Channel interface. You can implement iSCSI
interfaces instead of Fibre Channel or install InfiniBand HBAs.
iSCSI
Historically, iSCSI has not found its place in large-scale
implementations, perhaps because Ethernet was nowhere near as fast as Fibre
Channel. iSCSI over a Gigabit Ethernet (GbE) fabric is great for a few servers
accessing a small SAN, but it just doesn't have the juice for more than that.
However, with the advent of 10GbE, that's no longer the
case. You can team multiple 10GbE adapters in a file server node, and provide
data throughput rivaling that of Fibre Channel for a fraction of the cost. Four
10GbE adapters is about two-thirds the cost of a pair of 16Gb Fibre Channel
adapters.
InfiniBand
Alternatively, if you want to go the super-fast route, you can install InfiniBand HBAs in the file
servers. A pair of 40Gbit/sec InfiniBand HBAs is less than the cost of a single
16Gbit/sec Fibre Channel HBA, and a
54Gbit/sec InfiniBand adapter is cost-comparable to the Fibre Channel adapter,
but it gets you into the hundred-gigabit throughput range on each file services
cluster node.
Compute Nodes
On the compute node to file server link, you can team
gigabit adapters to provide multi-gigabit/sec throughput to each compute node,
or even install a 10GBit/sec Ethernet adapter and dedicate some or all of that
bandwidth for file services.
For virtualization hosts, you also have 40GBit/sec Ethernet
adapters in the same price range as Fibre Channel, and on the horizon we have
100GBit/sec Ethernet. The advantage here is not only in the cost savings of the
Fibre Channel adapter, but also in the freedom from expensive Ethernet switches
instead of Fibre Channel switches.
All of these technologies will benefit compute node
connectivity to the file servers, as well as present additional options for
connectivity from the file services cluster nodes to the SAN.
Scalability
Another advantage of using a file services cluster to
front-end the SAN is that the file services cluster can scale horizontally with
ease and minimal additional expense. For the cost of one additional connection
to the SAN, file services can be provided to dozens, perhaps hundreds, of end-users
via SMB v3.0.
In addition, there are no major initial implementation costs.
In fact, the methodology could be implemented with a single standalone file
server! From a practical perspective, however, a two-node cluster would be the
place to start for a production network.
The WS2012 file services cluster can scale to eight nodes,
providing a virtually unlimited SMB v3-based file access capacity for compute
nodes-including virtualization hosts and the dozens of virtual machines that
will run on those hosts.
Summary
While the title "Reshaping the Virtual Storage World" may be
a bit of hyperbole, there are certainly elements of truth to it. As memory and
compute capabilities have gotten cheaper, storage IOPS have increasingly become
a limiting factor for virtual environments. This is especially true as
companies have just started to consider moving high IOPS applications like
databases off of physical servers and onto virtual ones. This ability to
significantly increase storage throughput at a relatively low cost with a
software asset truly could start to reshape the virtual storage world.
###
About the Author
Lawrence Garvin is Head Geek and technical product marketing manager at
SolarWinds, a Microsoft Certified IT Professional (MCITP), and an eight-time
consecutive recipient of the Microsoft MVP award in recognition for his
contributions to the Microsoft TechNet WSUS forum. He has been working with
Microsoft Windows Server Update Services (WSUS) and Software Update Services
(SUS) since the release of SUS SP1 in 2003, and update management, generally,
since the creation of Windows Update in 1998. Prior to joining EminentWare (now
part of SolarWinds) in 2009, Lawrence offered Windows Server Update Services
expertise, including deployment, implementation, and troubleshooting advice to
companies worldwide as Principal/CTO of Onsite Technology Solutions.