So you are thinking of making the leap from legacy, static infrastructure – which makes this new era of heterogeneous computing really difficult to manage – to the cloud-like flexibility of composable architecture. Now what? In this new world which sounds almost too good to be true, where you can use software to add or subtract servers, GPUs, and FPGAs instead of physically installing or removing components from a server chassis, what are the gotchas? Let’s explore what features and functionality components matter, so you’re not left a casualty of the bleeding edge.
Composable infrastructure, with its ability to disaggregate all data center resources into pools of hardware resources, promises the agility and flexibility of the cloud at a fraction of the TCO – and with the security and control of on-prem deployments as a bonus.
The ability to have workloads configure servers on the fly without having to send someone to manually connect the devices to each other sounds like an IT manager’s dream come true: no need to determine workload requirements at server provisioning and configuring time; no need to purchase and procure the hardware components and wait weeks or months, especially in these times of parts shortages; no need to assemble, test and deploy the server.
All of this is accomplished via software and translates into one of the key benefits of composable disaggregated infrastructure (CDI): faster time to results. In addition, composability improves resource efficiency by increasing the utilization rate of your existing stranded resources, and it makes it feasible to build otherwise “impossible servers” – think 32 GPUs on one server. Everyone from Gartner analysts to the CEO of NVIDIA touts composable as “the future of the data center”.
What are the key functionalities to look at under the hood of the various vendor offerings then? Here is a quick checklist of questions to ask – a more detailed version can be downloaded here which you can print out and use to take your own notes as you make the rounds of the various options in the marketplace.
- Composes components in a disaggregated external chassis?
- Also composes those components physically inside the server?
- Composes at sub-device level – fractions of a device (e.g., MIG/SRIOV)?
- Even composes server-hosted storage or GPUs inside another server?
- Scales compute clusters solely over their native “language”, PCIe, without requiring the expense and administrative hassle of setting up and maintaining Ethernet or InfiniBand networks as well?
- Can use the memory from another server?
- Networks servers together over PCIe?
- Runs MPI over PCIe?
- A single PCIe network throughout the rack?
Software options and requirements:
- Requires proprietary software?
- Open platform with industry standard Redfish® APIs?
- Fully qualified with off-the-shelf composition and orchestration software?
- Simple to use if you prefer using Chef, Puppet, or Ansible scripts?
- Ability to daisy chain, leaf and spine, and cascade switches for scaling out?
- Ability to revitalize any existing server to add the latest and greatest technology – GPUs, storage, etc?
- Fully managed hardware environment?
As you can tell from this summary list, there are a lot of factors to take into consideration when investigating what is behind the slogans and the hype. Hopefully these questions will be useful in your search for the right fit between your data center needs and the vendors promises☺
For a more exhaustive list and a handy-dandy printout where you can jot down your own notes when comparing vendors, download this comparison tool. Happy comparison shopping!
Of course, if you are ready to test-drive the technology, schedule a demo.