We hear a lot about disaggregated and converged systems as hot technology trends that enterprises should adopt. But isn’t disaggregated the opposite of converged? How can that be? Let’s take a closer look at the terms to sort through their meanings and how they impact choices for IT organizations.
Network disaggregation allows choosing software independently of hardware and describes white-box (merchant silicon based) network switches running an independently offered network operating systems.
This is similar to x86 servers running Microsoft Windows, GNU/Linux perhaps along VMware vSphere or KVM hypervisors. We take server disaggregation for granted, but network switch disaggregation is new. Network operating systems include Big Switch Networks Switch Light (based on Open Network Linux, an Open Compute Project (OCP) project), Cumulus Networks Cumulus Linux or Pica8’s PicOS. Juniper’s Junos runs on its OCX1100 switch, an OCP-submitted design.
Hardware vendors for disaggregated network switches include white-box vendors such as Accton, Agema, Alpha Networks, Edge-Core, Penguin, and Quanta Cloud Technology. Traditional networking vendors providing "branded white boxes" include Dell Networking, which continues to sell its own FTOS network OS-based switches as well as disaggregated switches with multiple third-party operating systems. HP Networking resells Accton switches with Cumulus Linux alongside its traditional switches. Juniper has an open networking switch for Junos in addition to its traditional devices.
Converged systems are trickier to define. One popular definition has compute, storage and networking pre-configured in a turnkey manner. ESG, the firm I work for, uses the term “integrated computing platforms.” Hyperconverged infrastructure goes a step further by combining hardware and software into building blocks for cloud infrastructure. Let’s use the simple definition of an integrated platform where components are pre-configured for ease of deployment and operation.
These converged platforms are a reincarnation of sorts. Computer systems used to be tightly coupled become disaggregation became the trend, offering a choice of OSes and separation of disks to file servers or SAN systems. We now see a reemergence of closely coupled converged systems, but for new technical reasons.
You may view disaggregation as a new trend for network switches, but convergence as a server trend, and that settles the conflict. But this matter is not that easy resolved in the world of IT.
In Cisco’s UCS M-Series Modular (disaggregated) server, often under-utilized parts such as power and I/O, are aggregated and shared between different compute nodes, so this decoupling enables different parts to be independently upgraded. Since Cisco’s UCS servers may be part of a converged system (Cisco calls it integrated infrastructure), we see disaggregation within a converged system!
Mellanox disaggregates using RoCE (RDMA over Converged Ethernet) as a fast interconnect to separate CPU and memory from storage, and transmit frequently used data to SSDs and less critical data to SAS/SATA drives. The vendor also has multi-host disaggregation technology that takes multiple servers and connects to one NIC.
So disaggregation can mean different things: software/hardware choice, modular upgrades, physical separation of components, or sharing components depending on product or vendor. This leads to different benefits and use cases.
With such varied meanings, we can’t claim that disaggregated systems are always better than aggregated systems or that converged systems are better than non-converged. It’s not a simple matter of identifying your organization along a simple scale, such as number of network switches and using that to decide, for example, that a large number of devices requires disaggregation. It’s better to understand the business requirements before making purchase decisions, such as improving capex or opex, upgrade cycles and budgets, or choosing operating processes appropriate for the size of the infrastructure.
Using the example of networking switches, disaggregation provides vendor choice, is DevOps friendly and can lower capex if white-box switches are purchased in sufficient volumes. But the tradeoff is the increased effort needed to integrate the components and operate them.
Large cloud service providers or financial institutions possess the scale to offset this effort by amortizing the cost over large deployments. Cloud service providers such as Facebook and Google use disaggregated networks and have designed their own switch software and hardware adapted to their DevOps processes.
For mainstream enterprises, turnkey purchase and deployment, support and ease of operation is important. Aggregated systems simplify deployment and support, and leverage their IT staff’s current skills. Mainstream enterprises typically have network teams manage network devices, rather than treat network devices like another server in the DevOps manner as big cloud companies like Facebook do.
I recommend avoid using rigid definitions of “disaggregated as good,” or “unconverged is old-fashioned.” Turn your gaze away from the external messages and view your internal infrastructure needs instead. Although many IT organizations are surprisingly similar, there are still critical differences.
When considering adopting either disaggregated systems or converged infrastructure, examine your company's skill sets, need for deploying and managing the systems, tolerance for risk, projected budget and deployment scale, and then make an informed decision.