When it comes to compute power in the Telco and operator data center environment, low power consumption and high efficiency are ideal to support scale-out, cloud-based services and workloads. At the same time, there has been an ongoing, exponential volume increase in customers, data, and services, which is putting extreme pressure on service providers to increase server performance and productivity while still keeping their operational costs down.
SmartNIC platforms are enabling operators to achieve these goals through offload and acceleration—in order to deliver maximum performance for the support of efficient, scalable, programmable, and highly secure software-defined networking (SDN) and network functions virtualization (NFV) solutions.
In this interview, Niel Viljoen, CEO and founder of Netronome, discusses the new networking requirement in the data center, the failure of Moore’s Law, and a new wave of computing that he believes represents the future.
SDxCentral: Why is the demand for SmartNICs expected to accelerate so dramatically in the next few years?
Niel Viljoen: As you know, more and more networking is being done in software in the server, with virtualization and security being key drivers. Even while ignoring these issues, if you consider the speeds at which those servers are connected, the need for more networking has grown exponentially. Ten years ago, the average connection speed was about 1Gb/s; three years ago, it was 10Gb/s. Today’s interconnections average around 25Gb/s to 40Gb/s with some instances of 100 Gb/s happening now.
Because of this increase in speed, there is an order of magnitude more processing in the server that’s required to support the network. In addition, some security workloads such as DDoS have larger numbers of small packets which exacerbates the problem. At 100Gb/s, that’s 100 times more work for the processor than it was 10 years ago. Current general-purpose CPU or multicore SoC-based solutions cannot handle a 100Gb/s interface from a software perspective in an efficient way. Imagine playing the latest, greatest video game out there with 10-year-old CPU trying to handle all those graphics. This is the same issue: it just doesn’t work, and server productivity falls to zero because all resources are used to support the network.
The conventional wisdom has been that Moore’s Law will save us as more processing power becomes available without additional costs. But compute capability hasn’t scaled at the same rate as networking speeds have, and due to the demise of Moore’s law, this effect has been blunted. So the only solution is to add more “co-processors” ushering in the fourth wave of computing. This is typically characterized by parallel processing processors tightly coupled to standard CPUs focusing on networking, machine learning, search, etc.
You must have new server approaches to enable all of this. That’s why the movement towards app-specific accelerators, SmartNIC being one of these, is now growing enormously.
SDxCentral: Where do you see the market topping out?
NV: This is the future—SmartNICs will take over the market for networking in servers. The market will also expand beyond today’s addressable share, because people are now writing applications for the SmartNIC market to take advantage of the acceleration. This is analogous to the changing use of GPUs (graphics processors) where they were originally used as functional elements but new application areas emerged with the availability of programming tools such as CUDA from NVIDIA. This opened many new markets, self driving cars, AI, etc. In a similar way we will see new applications arise.
This is already happening in the SmartNIC space. Most of the hyper-scalers have deployed apps that leverage its capabilities. For instance, the more that virtual machines (VMs) are deployed, there’s a need to balance capacity to ensure performance. So, you need telemetry and monitoring of the applications running in those VMs. Often, the whole of such a management application ends up on a SmartNIC. Scalable denial of service mitigation in the face of enormous packet volumes is another use case. And the list goes on. Another use is where the SmartNIC becomes the control that the IaaS vendor has as they turn over the complete server to customers, the so called “bare metal” model that allows customers to load any legacy apps on servers in the cloud.
SDxCentral: Netronome recently announced that it is working with Qualcomm to make it possible for operators to deploy more VNFs and still maintain a decent throughput. Can you tell us more about this partnership?
NV: The Qualcomm Centriq 2400 can offload networking and data plane/path tasks to our Agilio SmartNICs. That allows the Qualcomm processor to focus on processing applications, enabling Telco and cloud operators to deliver more innovative services, distributed security, and visibility functions to a larger set of end customers while lowering TCO [total cost of ownership] to data center infrastructure operations. Efficient load balancing can assure that the large numbers of cores are used efficiently.
People that have specific workloads that are reasonably independent, or who are trying to optimize on IaaS for a lower cost than Amazon provides, will find that the integration between the CPU and the SmartNIC drives new price points and is much more efficient. We have already shown stunning performance figures.
SDxCentral: How will the growth of artificial intelligence (AI) and the cloud impact Netronome’s SmartNIC business?
NV: There are two aspects, firstly the CPUs will need more networking performance in larger and larger clusters. SmartNICs release the cores for the AI applications and potentially deliver lower and more deterministic latencies to the applications.
The second is where the SmartNICs can be used to aid AI applications on network traffic. This is already the fastest growing are in network security and application optimization. SmartNICs can extract all the features that machine learning applications in these areas require. SmartNICs can also contain machine learning engines implemented inside or as partner FPGAs.
In hyperscale cloud environments, this information extraction and analysis is extremely important, because that’s what customers are paying for. Add in AI to the mix, and what can be done—and what services can be built—becomes very exciting.