Over the last five years data centers have changed dramatically. Top cloud companies are moving away from converged infrastructure (CI) – where compute servers, storage and networking equipment are packaged in a pre-qualified turnkey solution – toward hyperconverged infrastructure (HCI).
HCI typically leverages commodity servers that are managed using software. Because they are software-defined, each server can handle a wider range of workloads. And unlike CI, HCI scales out: when more compute or storage is needed, data center operators simply add more nodes.
Now we are on the cusp of another sea change in technology — the transition from HCI to Composable Disaggregated Infrastructure (CDI). CDI frees data centers from the physical limitations of servers, allowing compute and storage to be housed in separate servers, interconnected by network elements.
There are two key elements to CDI:
- Composer software. By using software that offers flexible provisioning, orchestration, and automation, infrastructure resources can be composed of the exact elements required for a particular workload. Unlike HCI, the ratio of compute to storage resources is picked based on workload demands. When the job is finished, those resources are available for other workloads.
- Disaggregated hardware. Separating physical resources allows data center operators to choose best-in-class components, which can then scale independently.
Disaggregating compute and storage enables data centers to more easily pool these resources and provision them on demand as workload needs change. When every component can scale independently, data center operators can avoid the over-provisioning of compute or storage that often accompanies HCI. They can choose the best hardware for each component and maintain them on their own life cycles.
As computing elements inside data centers become more heterogeneous, incorporating specialized domain-specific silicon such as GPUs and FPGAs to accelerate application workloads, HCI servers become more difficult to support. The number of potential permutations increases dramatically, as do the management challenges. Because CDI doesn’t rely on a fixed configuration of compute and storage elements inside each box, there will be far fewer server variants to worry about. Provisioning, orchestration, and de-provisioning will be automated and instantaneous.
IDC predicts that adoption of CDI will grow at nearly 60 percent per year, reaching a market size of $3.4 billion in three years. Even then, it will account for less than 5 percent of data center infrastructure.
There are a few fundamental requirements for realizing the full potential of CDI. First, compute and storage resources must connect to a very high bandwidth, reliable and low-latency IP over Ethernet (IPoE) network fabric to ensure there is no consequential performance penalty across servers. Second, these compute, storage and GPU servers must be highly efficient; data-centric computations must be offloaded, so that the CPU is no longer the bottleneck. Third, applications running on CDI should perform at least as well as under HCI, without requiring changes to their programming model. Finally, CDI must be able to work at every scale, from a few dozen servers to thousands of racks.
Today, many expensive data center resources are stranded behind CPUs that are not designed to efficiently handle the huge volumes of data flowing through the network. As a result, these resources are starkly underutilized. This is why we believe that an entirely new type of silicon, the is required to enable the new paradigm of data-centric computing.
Freeing data centers from the constraints of the CPU will both greatly enhance resource utilization and reduce latency. And because the data center’s existing resources will be used more efficiently, operators should see a dramatic improvement in performance without any increase in operating expenses.
While data centers will continue to rely on HCI for years to come, we believe the future lies in more flexible and efficient composable solutions enabled by DPUs. Operators that aggressively invest in CDI technologies will quickly reap the benefits.
To learn more about composable disaggregated infrastructure, watch this.
Look for my next blog post to learn about a data center-friendly programming paradigm for DPUs.