BACKGROUND IMAGE: iSTOCK/GETTY IMAGES
Hyper-convergence represents the latest step in the evolution of data systems architecture. It combines the management...
of virtual servers, storage and the network to provide operators with a single, seamless view of a pool of computing resources.
Earlier computer systems consisted of individual servers, storage and networks. Converged systems represented the next step. These systems brought together compute and storage resources along with networking, but each component was still viewed as an individual element.
Hyper-converged systems build on these concepts. Network system management is simplified, since a single view eliminates the need for separate management consoles for compute, storage and the network. Compatibility issues disappear, since the entire system is typically delivered by the same vendor or system integrator.
Hyper-converged systems are available from specialized vendors, including Nutanix, Pivot3 and Stratoscale, and from large system vendors Cisco, Dell EMC and Hewlett Packard Enterprise. Systems are built on commodity server hardware and are often delivered preconfigured and ready for the customer to quickly install, start up and put into use. As overall computing load increases, additional systems can be added to form a cluster of systems; software integrates the individual systems to offer a unified view of the total set of resources.
The worldwide converged systems market, which includes hyper-converged systems, grew 4.6% year over year in the first quarter of 2017, according to IDC, topping $2.6 billion.
Network performance is critical
Hyper-converged systems place demands on the network beyond those placed by conventional systems. Cluster performance requires allocating sufficient network bandwidth, so it's important when designing a network to understand the level of traffic and the quality-of-service (QoS) requirements a hyper-converged cluster places on the network.
In addition to ensuring sufficient bandwidth, latency is a critical factor. It's important to select low-latency switches and build a flat network -- as opposed to other topologies -- designed with a minimal number of switches between applications and data.
Application performance depends on rapid access to data, but data required by an application may be scattered across multiple systems within the cluster. Unlike a traditional network design, hyper-converged clusters have no separate storage area network.
Disks within an individual system are direct-attached, but data on other systems in the cluster must be accessed via the same Ethernet network that carries application to application and management traffic. The management system may meet performance requirements by configuring virtual LANs with QoS settings that prioritize storage data over other data. But in some cases, adequate performance requires configuring dedicated Ethernet links for storage traffic only.
Supporting a variety of functions
Many hyper-converged systems support data replication and deduplication. Multiple -- often three copies of each database --are maintained on different network management systems within the cluster. Whenever an application makes a change in a database, that change must be propagated across the network. This traffic generates a constant load on the network, while deduplication adds traffic as redundant data is removed from multiple copies of the database.
The cluster balances loads across systems by moving virtual machines (VMs) from a heavily loaded system to a less loaded system. Unlike data replication, this traffic does not represent a constant traffic load. It occurs only when a VM moves, but the network must be designed with enough capacity to support ongoing data movement for other applications while moving a VM.
VM clones are created to support multiple users executing the same application. In some cases, a clone may execute in another VM on the same system, but when many clones are created, they will often be placed on systems across the cluster. This operation places a heavy, but short-term, load on the network. Snapshots capture the state of a VM, so it can be resumed at that point in case of later failure; copying snapshots across the network adds additional load.
Hyper-converged clusters must keep running even when a server or network link fails, so the network must have redundant links. VMs can be reconstructed and continue to execute by accessing snapshots and replicated data, but none of this will be possible if the only network link to a section of the network is down. Redundancy is also required for when a switch is taken offline for an upgrade, or when the cluster grows and a switch must be upgraded to support additional systems.
Network management system software must be flexible
The network component within your system management foundation must interface efficiently with VM hypervisors and storage management -- allocating and configuring data paths as VMs are created and terminated and when VMs are moved. Sometimes, it's necessary to make changes throughout the network to support a new application while continuing to provide required resources to other applications. All of these operations must take place without operator intervention, because it would not be possible to operate the cluster efficiently while depending on manual commands.
Additionally, network management system software must recover from system or link failures to keep the cluster operating. It must also provide accurate and timely reports of hardware failures and warnings about intermittent failures that may signal future trouble. Reports on processors, storage and network links show over- or underutilized resources. Trend reports make it possible to plan upgrades before cluster performance begins to degrade.
The advantages of hyper-converged systems -- reduced management costs, efficient use of resources and a smooth upgrade path -- are now widely recognized. With its increased use, more networks are expected to transition to the new architecture, and small clusters will grow into large clusters. As networks become more complex and applications with differing requirements are added, the benefits of hyper-convergence will become more apparent.