Tuesday, October 5, 2010

As discussed in our last article “What is Scale-Out Storage” this architecture solves many problems for the data center, the biggest being the need to scale performance and capacity in harmony with each other as the demands of the environment continue to grow. Another solution scale-out storage provides is the ability for near limitless growth. However, what scale-out storage really brings to the enterprise is a new way to mitigate the risk of running storage without traditional 2x capacity/performance buffers, and the impact of server virtualization.

The combination of tight budgets and server virtualization is creating the opportunity for storage operations to put the organization at risk by impacting the data center’s ability to respond to business needs. As stated earlier budget tightness has removed the ‘resource padding’ that used to be available. It’s realistic for most businesses even when dealing with these budgets to forecast and spend six months to a year in advance of what current storage needs are. The challenge is that most storage systems can’t follow that sort of growth curve, especially in terms of performance. Almost all of them have three to five year utilization models, meaning that the vendor will want the customer to predict their storage I/O demands three to five years into the future. Both of these are now impossible thanks to budget restrictions and the unpredictability of virtualized workloads. If the prediction is wrong, an upgrade, possibly a significant in scope, is required.

Most storage systems, especially in the entry level to mid-range enterprise, come with all the storage I/O performance installed upfront. There is no way to add more storage processing power and often there is no method for adding storage I/O channels. Storage capacity growth will only decrease the performance of the system as the storage processor becomes consumed with managing that capacity. Storage processing is particularly important as the environment grows. Larger environments mean more servers to connect to shared storage, more snapshots to manage, faster and more hard drives to manage and potentially the deployment of solid state disk drives. All of the above factors put more pressure on the storage processors to maintain the data flow to and from the storage sub-system.

Server virtualization makes matters worse as it becomes more prevalent in the data center. In the past, most of the servers attaching to shared storage did not have a sustained storage I/O need. There were lapses in demand and all that had to be ensured was that not too many servers had a sustained need at the same time. Server virtualization combines all of these spaced-out workloads onto a few servers. And while there are less physical servers to contend with, each has the potential to produce a steady stream of highly random storage I/O types. For example, it’s possible for a single server to have sequential and random I/O requests being sent to the storage system at the same time.

The combination of having to buy storage systems that were only meant to solve the coming year’s storage I/O needs plus the I/O unpredictability of the virtualized environment have forced many enterprises to slow down both their server and storage consolidation plans. After all, they must make sure that the consolidation efforts that are already underway will have the overhead to handle any peak demands in service. Ironically, their inability to further consolidate servers and storage is costing these organizations the very money they hoped to save by buying less capable storage.

When the storage I/O performance or storage capacity ‘wall’ is reached the organization is faced with two equally unattractive options. The first, and seemingly the most logical solution, is to buy a second storage device from the same vendor. However, even if this is the same exact type of storage and even if the vendor of that storage can provide an interface that will allow a storage manager to “see” both devices at the same time, actual management of the devices must almost always be done on an individual system basis. In other words, the common management window is really a monitoring window.

Often these additional devices, particularly in virtualized server environments, come from different vendors (to save costs) - basically buying from the cheapest supplier at the moment. The hope is that the server virtualization software will make it easier to manage these different platforms. Unfortunately, this is rarely the case. In almost every situation creation and management of LUNs, as well as other storage management tasks, is still a manual array-by-array process. In short, the storage management problem still remains, IT efficiency decreases and the organization suffers.

The other option is to upgrade the current storage system to one that has additional processing power and storage I/O channels. While this certainly keeps storage proliferation under control it’s an expensive option and one that will likely be repeated. While many vendors can now perform these upgrades and preserve the existing data set, most require or advise some sort of down time as the new storage controllers are put into place.

The alternative is to go with a scale-out storage system similar to what Isilon offers in their OneFS powered scale-out storage systems. Scale-out storage allows an IT organization to buy only for the current year or current quarter, the needed hardware to meet storage I/O and capacity demands. Then, as needs grow, the storage manager can easily add performance and capacity a node at a time. A scale-out storage system is designed from the outset to increase storage I/O and capacity without requiring (or even advising) downtime. The additional nodes connect into the architecture and are immediately recognized and used by the system. No additional work is required on the part of the storage manager.

Trying to plan which physical host is going to need the most storage I/O and when it’s going to need it, is almost impossible in the rapidly changing virtualized environment. Trying to isolate which physical servers need the most performance is a moving target, since virtual workloads can shift from physical server to physical server at any moment, often without the system administrator even knowing (thanks to distributed workload automation). Scale-out storage solves this problem by making sure that enough total bandwidth is available for the infrastructure, not just for particular physical machines.

Scale-out allows IT to run closer to the ‘resource edge’ without putting the organization at risk of application slow down or failure caused by storage. The scale-out storage model more closely matches the reality of today’s budget while it complements the server virtualization I/O requirements. Essentially, scale-out storage allows a high end enterprise-class storage system to be acquired in bite-sized budget chunks over time, instead in one huge investment upfront.

George Crump, Senior Analyst