As we start 2014, storage is changing. The familiar dual controller, scale up architecture is potentially being replaced by a new set of architectures. Storage professionals now have to consider storage systems that can scale out or converge themselves with the existing compute architecture. In this article we will help you determine if you should replace a traditional scale up architecture with one of these more modern storage incarnations.
Scale Up Lives
The legacy scale up, fixed controller architecture has served us well. And for many customers, both now and for the foreseeable future, that architecture will continue to serve them well. If a standard shared storage system can meet both your capacity and performance needs, both now and over the next five years, then there’s no need to change. These are simple, well understood systems that solve many of today’s performance challenges.
Scale out architectures are ideal for data centers that do expect to grow quickly. In the past these organizations had two choices in selecting a storage system. First they could over-buy, spending precious upfront capital on a system that they would hope to eventually grow into. Or second, they could buy to meet today’s need only, knowing that they would either have to replace that system or manage multiple systems.
Scale out architectures address this problem by allowing storage and capacity to be added in unison using clusters of dedicated storage servers or “nodes”. Each time a node is added to meet increased demand for capacity additional compute power comes along with it so that performance does not degrade.
Clustering, whether it’s a cluster of storage nodes or cluster of database servers, introduces latency. This latency is caused by two factors. First, the cluster itself requires inter-node communication to distribute data across the nodes of the cluster. Storage latency is also caused by the latency of the network, the time required for data to move through the network elements between the storage system and the connecting servers.
In the past the latency of the cluster was not noticeable compared to the latency of hard disk drives. However, the data centers, and the scale out storage systems in them, are moving to flash based media, which does expose the latency of a scale out architecture.
Faster networks, more node compute power and better cluster design can all help compensate for storage latency but it’s something that scale out system vendors need to design for. Also, each of these improvements, other than potentially better cluster design, can add cost to the scale out storage solution. SolidFire is a good example of a company that has developed better clustering code, and Isilon is an example of a company that has compensated with a faster inter-connect network (Infiniband) and more powerful compute resources.
Converged architectures offer an alternative to scale out storage by moving the storage function to the compute tier, which is increasingly virtualized via hypervisors like VMware. These systems typically aggregate storage from within each server and present it as a shared pool that all the servers can store data on. Their shared states make them ideal for clustered and hypervisor environments.
Some of these converged solutions have the ability to keep a copy of the virtual machine’s data on the server that VM is running on while replicating its data to the shared pool described above. With the data available locally all reads happen at in-server, transfer rates, but data is still protected and available if the VM gets migrated or the host server fails.
These architectures resolve a couple of issues that shared, scale out architectures are experiencing. First, they can eliminate both the cluster and network latency described above because the data is now on the host server that the VM is running on. It doesn’t have to wait for cluster synchronization to respond to data requests.
The second scale out issue that converged solutions resolve is better utilization of compute resources. One of the challenges that scale out architectures face as they expand capacity via additional nodes is the over availability of compute resources.
Converging the storage software onto the compute architecture allows for the compute resources to be better used while eliminating additional, unused compute found in the storage nodes. The result could be a significant capital savings.
Converged Architecture Implementations
Converged architectures can be delivered as turnkey systems from vendors like Nutanix, or they can be delivered as a software component from vendors like CompuVerde, OnApp, Simplivity, Maxta, and EMC’s ScaleIO. Certainly the turnkey approach has its advantages but many customers like the flexibility to select their own hardware, both servers and storage.
Selecting The Right Architecture
Potentially the biggest challenge for IT professionals is selecting which of these possible configurations is best. Our take is that if you can get by with a traditional scale up architecture then you should. They are simple and well understood. But this means it must be affordable upfront and scalable enough to meet your needs over the next five or so years.
A scale out storage system is ideal for capacity centric environments. This can also be a good choice when both capacity and performance demands are expected to grow rapidly. The most typical example of data centers that fit this requirement are online application providers and especially cloud storage providers.
Interestingly, converged architectures could appeal to both small and large enterprises as well as cloud providers. For the small enterprise they represent a very cost effective way to create a shared storage environment, leveraging the compute hardware they already have in place for virtualization. For large enterprises and cloud providers this technology can create a near perfect scaling model, as they add compute to satisfy client demands they are also increasing the appropriate storage resources.
Storage Swiss Take
IT professionals could see all of these architecture options as a negative. Sorting through the options and determining which one is best for your particular environment is going to take time. In many cases multiple solutions may fit from a functional perspective and the decision ends up being based on a subjective component, like supplier trust. But in the end choice is good for the data center, it allows you to drive storage capacity and performance down to the level that is appropriate for your business, while minimizing cost.