Victoria - Fotolia
When considering hyper-converged architectures, IT professionals need to decide what workloads the environment can support. The good news is that hyper-converged technologies are suited for most mainstream applications in the data center.
These are applications that need a modest amount of performance (sub-10,000 IOPS) and are not sensitive to a spike in latency. Applications that demand high performance or are sensitive to latency spikes often still require dedicated compute, storage and networks. Can hyper-converged storage vendors enhance their technology to satisfy all types of workloads?
A hyper-converged architecture creates a shared environment, so to some extent, it can only perform as well as its weakest link. While there are some tuning parameters in many platforms, it is almost impossible to ensure a specific performance characteristic to certain virtual machines (VMs). At best, a hyper-converged architecture can guarantee that a given application will get a higher percentage of the available compute, storage and network resources.
Areas of performance concern
Hyper-converged products have several performance weak points that should give IT professionals pause. Much of this stems from compute challenges.
The primary component of a hyper-converged system is the storage software. Many systems are essentially software-defined storage that has been optimized to run in a hypervisor cluster. These systems pool storage resources together and present them back to the VMs that the hypervisor manages. Some hyper-converged storage vendors include features like thin provisioning, deduplication and compression, as well as the required node management that a scale-out storage system needs. While these features can benefit the IT professional, they put an additional load on the infrastructure -- especially the compute component, which supports applications and VMs in addition to storage software.
Hyper-converged storage vendor promises
Many hyper-converged storage vendors claim that CPU performance will never be a problem because hyper-converged architectures scale out. However, the CPU resource is not aggregated like storage capacity is. Because the storage within a node is aggregated with storage from the other nodes into a single virtual pool, specific tasks need to be assigned to specific cores in each node in the cluster. Some of those cores must be assigned to the storage software processes. If a particular node runs out of cores, or processing power, it cannot offload its storage management task to another node in the cluster.
The most obvious workaround for this problem is to overcompensate and buy more compute resources than needed. But overcompensating by purchasing extra CPU drives up the price of the hyper-converged system, and as the environment continues to scale, the extra CPU is at risk of becoming the bottleneck.
Some hyper-converged storage vendors offer dedicated storage nodes. These are the only nodes in the cluster that are responsible for processing storage I/O and the storage software. Assuming enough of these nodes are bought, this should resolve the problem. Of course it does so at additional expense and, to some extent, breaks the concept of a converged solution.
The other alternative is to place dedicated CPU power into nodes that are designed specifically to process CPU-heavy storage operations like deduplication and compression. The advantage of this approach is that it maintains the integrity of the hyper-converged concept, provides a more consistent performance profile and does so with minimal additional costs.
Compute performance is a critical component to hyper-converged architecture, and it is also hard to guarantee to an application. For organizations with applications that require either high performance or very specific performance, IT professionals should look for platforms that can offload or dedicate CPU performance to storage I/O intensive operations.
Hyper-converged vendors offer different features
Weigh your hyper-converged architecture options
Features to look for with hyper-converged systems