Andy Bechtolsheim (Arista Co-Founder) – A discussion with Arista’s Andy Bechtolsheim (Apr 2012)


Chapters

00:00:00 Software-Enabled Functions and Extensions in OpenFlow
00:02:52 Challenges and Obstacles in Network Architecture Innovation
00:08:38 Centralized Management in Cloud Networks
00:10:51 Networking Technology Scaling and Evolution
00:13:59 Network Complexity Solutions: VXLAN and NVGRE
00:17:45 Custom Hardware vs. Programmable Network Gear
00:22:12 The Rise of FPGAs in High-Speed Trading
00:24:15 Software-Defined Networking: Challenges and Opportunities

Abstract

Revolutionizing Network Infrastructure: Arista’s Pioneering Approach in Multi-Node Systems, OpenFlow, and SDN

A New Era in Network Management: Arista’s Innovative Multi-Node System

In the field of network infrastructure, Arista Networks stands as a vanguard, particularly with its groundbreaking extension of the EOS (Extensible Operating System) model to multi-node systems. OpenSwitch’s EOS model of stateless agents and centralized state management in a single database seamlessly extends to multi-node systems. This simplifies and enhances features like hitlist upgrades, which have traditionally been challenging to implement. The model also offers opportunities for expansion into wider multi-chassis or multi-node settings.

The switch architecture is highly scalable, with some switches supporting thousands of nodes and managing multiple petabytes of traffic. Arista’s success in deploying scalable networking solutions in demanding environments is a testament to its effectiveness.

The OpenFlow Dilemma: Balancing Demand and Practicality

While OpenFlow has garnered significant interest for centralized switch management, Bechtolsheim notes its challenges, particularly in replacing traditional routing with centralized agents. The resilience required in node failure scenarios remains a critical consideration, tempering the enthusiasm for a complete shift towards OpenFlow. Nonetheless, Arista’s commitment to software-defined networks (SDNs) shines through, with a focus on leveraging software-enabled functions to enhance network capabilities.

However, OpenFlow has not met initial expectations, and flow-based network management faces scalability and practicality challenges. Networks evolve slowly due to factors such as financial investment, reluctance to change, and protocol maturity. Successful network technologies like ECMP and leaf-spine architectures have emerged based on proven and reliable approaches.

Challenges and Alternatives to OpenFlow

OpenFlow, despite its potential, faces hurdles in widespread production deployment. Issues like scalability and integration with existing network infrastructures, which aren’t typically managed on a flow basis, hinder its adoption. Bechtolsheim introduces ECMP (Equal-Cost Multi-Pathing) as a practical, scalable alternative. This approach, eschewing the need for centralized controllers, optimizes traffic distribution and reliability.

Complexity in networking involves multiple layers, introducing unnecessary complexity for layer two and layer three. VXLAN and NVGRE are standards that solve the problem of tunneling layer two over layer three. Ken Duda, Arista’s CTO, played a key role in developing these standards. Standard space protocols, like VXLAN and NVGRE, will succeed in the market due to interoperability between multiple vendors. Current silicon lacks hardware termination for VXLAN in the switch. Arista is considering implementing hardware termination in future silicon or FPGAs.

The Slow Evolution of Networking Protocols and the Rise of Virtualization

Bechtolsheim highlights the sluggish pace of change in networking protocols, exemplified by the protracted transition to IPv6. This inertia is attributed to substantial investments in existing infrastructure and a general reluctance to disturb stable systems. However, virtualization emerges as a transformative force, offering enhanced flexibility and resource utilization, thereby propelling new technological adoptions and architectures.

Advancements in Data Center Networking: Leaf-Spine Architecture

The shift from traditional three-layer architecture to leaf-spine design signifies a major advancement in data center networking. This topology, coupled with ECMP and layer two MLEC in the spine, optimizes network resource utilization, ensuring efficient load balancing and failover mechanisms.

Leaf-spine architecture with ECMP and layer two MLEC in the spine has become the standard deployment model for large cloud data centers. Arista focuses on data center networks within service providers, known as service provider customers. They do not offer the routing tables or features of full-fledged routers. Centralized management is a key goal in cloud environments to reduce OPEX and the number of network engineers required. Arista’s approach involves robust and trusted implementations to ensure reliability in the absence of maintenance windows.

Service Providers and Cloud Business Integration

Service providers are increasingly delving into cloud services, building expansive data centers within their networks. Arista, focusing on these data center networks, champions simpler and more robust solutions over traditional models. The key lies in centralized management, essential for cloud networks, aiming to slash operational expenses and reduce network engineer requirements. This approach, alongside the uniform deployment and standardization of cloud networks, offers pronounced efficiency and scalability benefits.

Arista’s Technological Edge: Scalability, Performance, and Compatibility

Arista’s switches are a testament to scalability, supporting extensive ECMP with large switches managing thousands of nodes. Regular silicon upgrades enhance performance, with features like increased bandwidth and larger tables. Importantly, Arista’s software supports various chip architectures while maintaining a single binary release, simplifying system administration.

Arista’s success in deploying scalable networking solutions in demanding environments is a testament to its effectiveness. Regular technology improvements are driven by Arista’s reliance on various chip vendors’ roadmaps. Silicon upgrades every 18 months to two years enhance performance, add ports, and increase memory and table sizes. Arista’s software supports diverse chip architectures with a single release, simplifying system administration by eliminating the need for separate release trains or version tracking. Arista anticipates reduced costs for high-speed optics, such as 40-gigabit and high-gauge optics, in the coming years. This development opens up opportunities for performance upgrades without extensive network changes.

The Complexities and Innovations in Networking

Bechtolsheim brings attention to the complexities of traditional networking, such as the layered VLAN approach. Innovations like VXLAN and NVGRE, tunneling protocols that extend layer 2 networks across data centers, offer solutions to these complexities. The potential of FPGA implementation in future chips further underscores Arista’s innovative approach to networking, providing flexibility and reducing vendor dependency.

FPGA: A Game-Changer in High-Speed Trading

In the high-speed trading arena, FPGAs (Field-Programmable Gate Arrays) offer remarkable advantages. Their ability to execute critical functions faster than software implementations, with sub-microsecond processing speeds, significantly reduces latency. This capability is especially beneficial in environments like Wall Street, where every microsecond counts. However, their suitability is confined to simpler applications, as complex functions like database management remain beyond their scope.

Faster Order Execution and Competitive Advantage:

– FPGAs provide sub-microsecond execution speeds by implementing logic close to the hardware level in the network switch.

– They offer a competitive edge by enabling faster decision-making and execution compared to software solutions.

Efficient Hardware Functions and Application Suitability:

– FPGAs handle functions traditionally handled by software, such as risk analysis and monitoring, efficiently.

– They are ideal for applications with a small number of gates or clocks.

– However, complex applications like Oracle databases are not well-suited due to increased coding complexity and inability to match the switch’s packet rate.

The Uncertain Path of Software-Defined Networking

The evolution of SDN is still shrouded in uncertainty, with limited adoption due to unique network requirements. Discussions around SDN are now centered on understanding its true use cases and exploring alternative solutions. Arista’s approach to SDN involves integrating it with their EOS and Cloud Vision platforms, enhancing network functionality without compromising on stability and operability.

Arista’s Practical Approach to SDN and Complexity Avoidance:

– Arista focuses on solving customer problems rather than blindly implementing SDN standards.

– The company highlights the challenge of making SDN standards useful and scalable.

Mapping SDN to Existing Platforms:

– Arista will map SDN use cases to its EOS platform and Cloud Vision.

– Cloud Vision allows multiple switches on a network to perform tasks beyond a single switch’s capabilities.

Arista’s Balanced Approach in Network Evolution

In conclusion, Arista Networks stands at the forefront of network technology evolution, adeptly balancing innovation with practicality. From extending EOS to multi-node systems to exploring the potential of FPGAs in high-speed environments, Arista demonstrates a commitment to advancing network infrastructure while maintaining the stability and reliability essential in production networks. Their approach, focusing on incremental upgrades and compatibility, paves the way for a more efficient, flexible, and robust networking future.


Notes by: ZeusZettabyte