?
Data Center Server Infrastructure
8 minute read

Rack-Scale Composable Infrastructure With Intel® Optane™ SSDs

Get an overview on scaling Intel® Optane™ High-Performance Storage with FabreX™ Next-gen CDI.

In This Article

One big challenge to managing data today is not just its vast, relentless growth — it’s also the fact that data are always on the move, an elusive target that must be processed, organized, analyzed and stored. With the surge of artificial intelligence (AI) and real-time processing of big data, speed of access and processing is becoming more crucial.

copy link

Data continuously in motion demands extreme performance

Some in IT speak of data existing at rest in “lakes” when a more accurate analogy would be “data fountains,” to communicate the kinetic quality of actionable data always on the move from various sources such as 5G, Wi-Fi 6 and edge computing. Besides velocity, actionable data are characterized by complex variety, from billions of 4-byte mini-files to enormous data structures.

Data and the information about the data (i.e., metadata) yield deep insights and intrinsic value in the overall decision-making processes. Whenever we discuss data that must be enduring, we couple the requirements and security of metadata accordingly. When files move, data and metadata move. The crucial discussion here is about data de-risking, and once the metadata is gone, the data are gone, too.

This roiling tempest of data can overwhelm the static nature of traditional data centers, with demands for ever-greater speed, performance and reliability. Organizations that can overcome these challenges stand a greater chance of winning the data struggle. But it can’t be done with yesteryear’s hard disks — only with high-performance solid-state storage combined with dynamic resource provisioning and sharing.

The explosion of computing capability built into modern CPUs, GPUs, FPGAs and SmartNICs has not been accompanied by a commensurate performance scaling of the networks over which the data generated by these devices must travel. As a result, network storage bottlenecks have appeared, which Intel Optane SSDs can help eliminate. In addition, WWT technology partner GigaIO further helps blast through those bottlenecks by reinventing modern data center architecture with FabreX — the world’s only all-PCI Express (PCIe) enterprise-class, open-standards, Rack-Scale Composable Infrastructure solution. FabreX unleashes the performance of Intel Optane SSDs throughout.

copy link

Why Rack-Scale Composable Infrastructure is the next-gen data center

As mentioned earlier, legacy data center infrastructures weren’t designed for modern workflows. Today’s solutions must integrate compute, storage and other communication I/O into a single-system cluster fabric, scaling resources across the cluster as needed. The solutions should free resources from their silos to be shared with other network users who draw from these resource pools through a composable disaggregated infrastructure (CDI) that maximizes usage and improves agility.

The Rack-Scale Composable Infrastructure powered by the FabreX hyper-performance network is a fundamentally new architectural paradigm that allows data center resources to be scaled up and out as needed — easily, rapidly and reliably — for maximum utilization and lower TCO.

The FabreX hyper-performance network fabric is uniquely able to achieve complete disaggregation and composition of all resources in a server rack — because it does so over PCIe, versus having to manage multiple Ethernet or InfiniBand networks within the rack, thus lowering or eliminating cost, complexity and latency.

copy link

What’s unique about Intel® Optane™ storage 

For those unfamiliar, Intel Optane SSD all-flash drives constitute a new tier in the memory and storage hierarchy, offering the most valuable characteristics of both. These rack-scalable SSDs deliver high throughput for breakthrough price/performance, high availability, low latency, greater responsiveness under load, predictably fast service and increased endurance.

Intel Optane SSDs help eliminate data center storage bottlenecks and allow superior, more affordable data realization. Ideal for accelerating applications, they reduce transaction costs for latency-sensitive workflows and improve overall data center total cost of ownership (TCO). Crucially, Intel Optane SSDs (and Intel® Optane™ persistent memory) are also where the all-important metadata should be classified and stored.

The key trade-off in storage is cost versus latency — Intel Optane SSDs provide the lowest latency in the continuum, especially in the caching or tiering of data. But for that latency advantage to be fully realized, the SSDs need to be direct-attached inside the server which limits their shareability and scaling potential — that is, until FabreX.

An important feature of Intel Optane SSDs is that they work with all PCIe 3.0 and now 4.0 slots. As such, they can be disaggregated and composed with non-volatile memory over fabrics (NVMe-oF). Latency reduction is growing as the crucial design element across the modern data center. Intel Optane SSDs, coupled with FabreX that securely operates with native PCIe, provide the lowest latency design possible, even with legacy networking in the data center.

copy link

Scaling, securing and sharing Intel Optane SSDs with Rack-Scale Composable Infrastructure

The key advantage of Intel Optane SSDs is low latency and high performance, yet these get partially or completely occulted when Intel Optane SSDs are networked over a legacy network with inherently higher latency and lower performance than PCIe.

Today, Intel Optane SSDs are constrained within a physical server, with legacy networks too slow to enable sharing without a latency and performance penalty. An all-PCIe FabreX architecture instead unleashes the core benefits of Intel Optane SSDs outside the node or the server, making it scalable and shareable across servers without a performance hit.

Because FabreX can communicate server to server over PCIe, the language all rack resources speak natively, it is unique in its ability to realize scalable Intel Optane SSD performance with significantly lower latency than other options for NVMe. Combining Intel Optane SSD with FabreX creates a synergistic effect for the highest performance scalable storage solution.

Without FabreX, Intel Optane SSDs are available only to the server or in a “just a bunch of flash” (JBOF) enclosure over a legacy network where its latency advantages are consumed by the network. Composable disaggregated infrastructure allows for a variety of placements with no degradation of performance or latency, as long as resources are communicating in their native language, PCIe.

In terms of security, indiscriminate hacking and malware can proliferate via open networks like Ethernet. Data traveling on a PCIe-only network never stops, hence eliminating the possibility of a security breach as the data are briefly stored in an HBA or HCA adapter with other legacy transports. Each touchpoint in a NIC is another opportunity for your data to risk being compromised or corrupted. In comparison, data traveling directly and exclusively over PCIe never stops. Hence, deploying the PCIe-native FabreX hyper-performance network provides an essential tool for your data center network’s security.

copy link

Performance testing

GigaIO recently conducted a series of latency tests of high-performance storage connected to FabreX in four configurations to demonstrate the performance.

Converged: on-board, physically inside the server
Figure 1. Converged: on-board, physically inside the server
Composed: direct composition, composed to the server over FabreX
Figure 2. Composed: direct composition, composed to the server over FabreX
NVMe-oF over FabreX
Figure 3. NVMe-oF over FabreX
NVMe-oF over InfiniBand
Figure 4. NVMe-oF over InfiniBand

Researchers found no discernible difference between Intel Optane SSD placement inside the physical server or composed to the server from outside the node — a mere 9 microseconds of latency in either case.

The complete software and hardware stack for NVMe-oF running over FabreX took just 3 microseconds longer — 12 microseconds. But when composed over InfiniBand, the latency increased by a factor of 3X to 32 microseconds, as expected because of the added InfiniBand translations. This increased latency completely eliminates the performance benefits of Intel Optane SSDs and renders sharing and scaling impossible.

Performance summary table
Figure 5. Performance summary table

The conclusion: IT organizations can finally network and share the power of Intel Optane SSDs across servers for more users to enjoy their latency advantages, as long as they do so over a native PCIe fabric like FabreX.

copy link

Unleash the power of Intel Optane SSDs with FabreX

Data is rapidly growing and continuously in motion, and traditional data centers can no longer keep up. Intel Optane SSDs, a breakthrough innovation in the memory/storage hierarchy that dramatically increases throughput to eliminate I/O bottlenecks, reduce latency and deliver excellent responsiveness. But until the new disaggregated, composable solutions from WWT partner GigaIO, it was difficult to scale Intel Optane storage beyond one node. 

It’s in the GigaIO FabreX environment that Intel Optane SSDs can finally be networked and reach their full resource-sharing potential — combining the best of both worlds for a synergistic effect. WWT can show you how to integrate these technologies to achieve optimum price-performance in your modern data center.

Learn more

GigaIO has redefined the possibilities for high-performance storage in composable infrastructure with its FabreX hyper-performance network fabric, coupled with high-throughput Intel Optane storage and persistent memory.

Learn more about each of these technologies: