Open Atomic Ethernet

Project Leads

Sahas Munamala, Paul Borrill

Ethernet has been the cornerstone of networking for over five decades, adapting to changing technologies, applications, and economic realities. New data center and edge opportunities—such as directly linking chiplet modules in ultra-low-cost meshes, achieving extreme low latency by executing application actions at hardware speed in the network interface, and applying these techniques to distributed updates of persistent storage or to distributed, application-level transactions-push beyond the boundaries Ethernet established in an earlier era. These needs cannot be met simply by extending Ethernet’s historical capabilities.

Open Atomic Ethernet will look ahead to the next 50 years of Ethernet networking. It will focus on ways to achieve higher performance, guaranteed service, lower latency, fewer errors, and greater scalability.

Scope

Align what the network does with what the distributed application needs to accomplish. Build a world where the network is not allowed to create a mess the application must discover and then clean up, but rather must either ensure both sender and receiver (user space software endpoints in the distributed application) know a message was successfully sent and received, or only the sender "instantly" knows it failed, and no other node or part of the network knows that message ever existed.

This entanglement between sender and receiver, with no library, software, or hardware permitted to disturb that entanglement, is a fundamental change from the "throw it over the wall and hope it gets there" of the last 50 years. From a different perspective, a network message becomes a transaction which exhibits ACID properties: atomicity, isolation, and durability directly, and consistency because that message exists only between endpoints and is not visible to any other observer.

• Rethink Ethernet’s core primitives: frame structure, error recovery, and flow control, in the context of Shannon Information Theory.
• Introduce API’s for programmable minimal compute capability at the NIC level to enable reversibility and support application-specific protocol behavior.
• Address the trade-offs between signals, noise and energy dissipation, building upon principles from Quantum Thermodynamics.
• New Fundamental multi-phase primitives: build state machines into the link itself to enable Network assisted transactions.
• Create a protocol framework that is as simple as Ethernet’s original design but scalable to today’s and tomorrow’s use cases, including chiplet interconnects and AI accelerators.

OCP TAP - Open Atomic Ethernet Project Calendar

The calendar displayed here is updated nightly from the project's Groups.io Calendar.