Hyper-convergence is hot. It's a pervasive topic in data centers these days that has intrigued storage and other IT managers, but has also generated a lot of confusion among users -- confusion that may be partly attributable to hyper-converged vendors and their aggressive marketing.
The hyper-converged appliance (HCA) product category is large and very active with new vendors joining and leaving the market, which can add to the confusion. In the following, we describe what hyper-convergence technology really is, and show how it's not necessarily synonymous with hyper-converged appliances and why that's an important distinction to make. We also discuss the differences among various HCA offerings and what those differences mean in the course of an evaluation of hyper-converged products.
Converged vs. hyper-converged infrastructure
To properly understand hyper-convergence, we need to start with converged infrastructures, the rack-based packages that combine existing storage, server and networking components. These systems provide either a complete storage and compute infrastructure at the rack level by selling commercially available components bundled as a single SKU (such as VCE's Vblocks) or a reference architecture from which users or integrators can assemble a complete system (such as the Cisco/NetApp FlexPods).
The primary value of a converged system for customers is a reduction in deployment time. Convergence essentially takes the integration out of building compute infrastructures, simplifying the implementation of off-the-shelf components, or it might just remove the design step as in the case of reference architectures, which still require some assembly. The scale of the typical product is fairly large as well, in terms of both physical size and capacity, with many converged infrastructures taking up most of a data center rack.
While converged systems bundle rack-level components into an integrated offering, hyper-converged infrastructures combine the compute, storage and networking functions into the same chassis. These infrastructures can be assembled from independent hardware and software components by users or integrators, or companies can buy a hyper-converged appliance -- a turnkey, single-vendor product that includes comprehensive support.
Inside the hyper-converged appliance
The hyper-converged technology category has become quite popular and very crowded as there are multiple ways to create a product that marries storage, compute and networking functions (See: "The Open Storage Platform"). For our purposes, we'll focus on hyper-converged appliances, the all-in-one products currently being sold by a couple dozen vendors.
To be classified as an HCA, a product must meet the following criteria:
- Hardware and software. While software is always at the core, HCAs aren't software-only products.
- Single-vendor product. HCAs should have one part number, require one purchase order and provide "one throat to choke" support; an HCA vendor must provide an end-to-end platform.
- Comprehensive management. An HCA is a turnkey product that, at a minimum, manages and simplifies the setup and initial configuration of the product and, ideally, provides simpler operation.
- Storage federation. There must be a software layer that virtualizes physical storage capacity in each node and federates those nodes into a common management point.
- Hypervisor. HCAs include an embedded hypervisor for running compute workloads and, typically, to run the VM-based management and the storage software.
EVO:RAIL hyper-converged appliances
There are two general categories of HCA products: those using VMWare's EVO:RAIL software and those that use proprietary software. VMware is the most common hypervisor platform in corporate data centers, and it enjoys this same advantage in the HCA market as the vast majority of hyper-converged systems support VMware. The company developed a virtual SAN product (VSAN) to provide an integrated scale-out storage product for VMware users. EVO:RAIL is a combination of technologies that leverages the proliferation of the ESX hypervisor to create a hyper-converged architecture that's easy for existing hardware vendors to use to create HCAs by bundling the software with their hardware.
EVO:RAIL uses the vCenter Server database, effectively integrating EVO:RAIL management into vCenter Server and simplifying HCA operations for users already familiar with VMware. This integration also means that changes made in vCenter Server will be reflected in EVO:RAIL and vice versa. And, instead of being run as a VM, VSAN is incorporated into the hypervisor kernel, potentially making the storage process more efficient.
There are currently eight qualified EVO:RAIL hardware partners: Dell, EMC, Fujitsu, Hitachi, NetApp, SuperMicro, Inspur and NetOne (the first six listed here are covered in Evaluator Group's hyper-converged appliance comparison research). EVO:RAIL uses a fixed configuration for CPUs, memory and storage resources that can be implemented by their hardware partners. With very few exceptions, EVO:RAIL-based offerings follow this configuration:
- Dual CPUs with a total of 12 or 16 cores
- Up to 256 GB of memory
- 4 to 16 node clusters using a four-node server chassis
- Up to 3.6 TB of hard disk capacity and 480 GB of flash storage capacity per node
An EVO:RAIL product allows users to get an HCA from an established company, since most of VMware's partners are traditional storage vendors. Also, because it's tightly integrated with VMware, it's an attractive choice for VMware users. Compared with the other HCA products, EVO:RAIL has limited configuration options and somewhat smaller scalability, although a 16-node configuration can support a large percentage of the typical HCA use cases. VMware is the only hypervisor supported and iSCSI is the lone storage protocol. In terms of storage features and functionality, EVO:RAIL doesn't support remote replication, deduplication, compression or multi-tenancy. This adds up to very little product differentiation for hardware vendors, with a few exceptions.
On the software side, Dell has recently partnered with Nexenta, adding NexentaConnect software to its EVO:RAIL offering. This brings inline deduplication and compression as well as NFS and SMB file protocols to the block-based iSCSI supported by VSAN. On the hardware side, SuperMicro almost doubles the standard EVO:RAIL per-node compute and storage resource limits and increases the maximum cluster size to 32 nodes.
Proprietary hyper-converged appliances
The most popular HCA products use proprietary software packages that run on industry-standard server hardware, but are sold as pre-configured modules. While they all use Intel x86-based servers, most are customized to look like purpose-built appliances. Hyper-converged appliance vendors who fit this category include Atlantis Computing, Dell (using Nutanix software), HPE, Maxta, Nutanix, Scale Computing and SimpliVity, among others. Given the activity in this dynamic product space, you can expect several more vendors to join this list in the coming year.
At the core of these systems is their storage software that provides the abstraction and federation layer required to pool physical storage resources from multiple server chassis and provide a centralized management function. While not specifically required, the vast majority of HCA products use a scale-out topology. Most systems support iSCSI, but two -- Nutanix/Dell and Scale Computing -- also provide file system protocols and Atlantis Computing supports NFS exclusively.
The Open Storage Platform for do-it-yourself hyper-convergence
Buying an appliance isn't the only way to get the benefits of hyper-convergence. Many of the vendors of software-defined storage are selling those products in conjunction with the appropriate hardware components and enabling users to create their own hyper-converged infrastructures. While these aren't hyper-converged appliances, based on the criteria we used for this article, they do offer a way for users to get an economical, highly scalable storage system that can support a compute function if desired.
These "roll your own" systems can be used in many of the same environments as HCAs, but they're not turnkey products. On the positive side, they offer more flexibility and often cost less. But they require integration by users or a third-party company, and can create support confusion between hardware and software vendors.
One example is VMware's VSAN Ready Nodes, which are validated configurations from qualified hardware OEMs that support VSAN.
Many models of proprietary HCAs
One of the biggest differences between EVO:RAIL-based HCAs and their proprietary counterparts is the number and variety of configurations available. Since HCAs by definition are pre-configured appliances, product flexibility is achieved by adding new models. Where EVO:RAIL offers basically two configurations, some HCA vendors have a half dozen or more models to choose from. This allows them to create products with different combinations of compute, memory and storage resources, and different scalability profiles. It can also add to the complexity of the purchase decision.
For example, proprietary systems routinely provide 24 CPU cores and over 500 GB of RAM. Cluster configurations typically start at three nodes, and several have no specified node limits at the top end. Per-node storage capacities are routinely greater than 20 TB for hard drives and well over 2 TB of flash for many models, more than four times the typical EVO:RAIL configuration. While the vast majority of storage configurations are hybrid, two vendors -- Nutanix and Dell (using Nutanix software) -- offer all-flash configurations. Atlantis Computing offers only flash products and Scale Computing offers only hard disk storage.
Scaling HCAs up and down
Hyper-converged appliances have enjoyed significant success for remote office and branch office (ROBO) deployments as well as in smaller companies. Besides ease of setup and operation, an important requirement for those environments is the ability to start small. For example, SimpliVity supports single-node clusters, although two nodes are recommended, and other vendors are expected to announce clusters of fewer than three nodes soon.
On the scale-up side, most HCAs simply add storage-heavy hardware configurations, but there are exceptions. HPE allows its HCA cluster to connect to its existing scale-out storage system (StoreVirtual VSA) and Maxta supports external, direct-attached storage (DAS) capacity.
Hyper-converged appliance features
The exact feature sets of hyper-converged appliances vary by product, of course, but common features for proprietary HCAs include local and remote replication, stretched- or metro-cluster support, deduplication, thin provisioning, compression, encryption, snapshots, clones, load balancing, quality of service (QoS), cloud-based backup and recovery, and WAN optimization. Most products support VMware and have integrated at least some management and control with vCenter. In addition, most support VMware advanced features, such as vSphere Distributed Resource Scheduler (DRS), vSphere High Availability (HA), vSphere Storage APIs - Array Integration (VAAI), and so on.
How the hyper-converged appliance roadmap is shaping up
The hyper-converged appliance product segment will continue to grow in the coming year, with more vendors joining the market. You should also expect some vendors to exit, from both the EVO:RAIL and the proprietary HCA categories. For example, HPE stopped selling its EVO:RAIL product in 2015 to focus on its proprietary technology, and a startup or two appear to be struggling to stay in business. Storage scalability will be an important characteristic with more vendors introducing storage-only node options and with more products supporting external storage systems or devices. At the same time, expect to see more two-node configurations being released to address the ROBO and small office use cases.
It's also likely that EVO:RAIL vendors will strive to differentiate their products, with larger capacity configurations and additional features. The wide variety of models in the proprietary category will continue as vendors cast about for new configurations to address new markets. That said; don't look for HCAs to expand into the traditional data center, as much as the HCA vendors might like that to happen. The cost and inflexibility of turnkey appliances make an HCA a less than ideal alternative in those environments. But new use cases will arise, such as private and hybrid clouds, and special projects like Hadoop and big data analytics that drive the purchase of new infrastructures. HCAs may prove to be appropriate for some of those use cases, but the Open Storage Platform may provide more compelling storage platforms for those scenarios.
Hyper-converged appliances benefit from reference architectures
Converged vs. hyper-converged appliances
A closer look at the hyper-converged architecture market