The NetDev 0x18 Conference, scheduled for July 15-19, promises to be a hotbed of innovation, with several exciting presentations lined up. Three papers, in particular, are expected to be game-changers in the world of AI networks and protocols and have a close connection to our work on the Infrastructure Data-Plane Function (IDPF) spec and driver.
The first paper, "The Future of AI Networks: Advancing TCP with Device Memory and Collective Communication," focuses on the evolution of TCP to support Collective Communication (CC) semantics. This initiative, originally introduced as Devmem TCP by Google, is a game changer in the AI network landscape. The paper focuses on Device Memory TCP network performance enhancements, similar to GPU Direct RDMA, for AI workloads. The authors will share their implementation strategies and provide updates on their progress, along with detailed performance data and future collaboration topics.
The second paper, "PSP Crypto Protocol and HW Offload," presents PSP, a security protocol that incorporates lessons from experiences with currently deployed L2, L3, and L4 crypto protocols. Developed and implemented by Google, PSP is engineered for scalability and avoids storing an explicit Security Association (SA) for incoming traffic. Instead, it dynamically generates the ingress key on the NIC using a device master key and the SA Security Parameter Index (SPI) found within the packet. This paper is a preview of a new capability that will be integrated in the IDPF spec and driver in the future.
The third paper, "Introduction to Falcon Reliable Transport," explains Falcon, a hardware-offloaded reliable transport that works with RDMA to solve RoCEv2 issues. Introduced by Google, Falcon is designed to meet the demands of workloads that require high burst bandwidth, high message rates, and low latency. The paper will detail the connection setup, key exchange, and integration with the rdma-core/ibverbs framework, along with proposed extensions to support advanced features. Falcon Reliable Transport leverages the newly defined interface for iRDMA from within IDPF.
Collectively, these papers signify a leap forward for AI networks, offering cutting-edge solutions to the complexities of network communication protocols, security, and reliable transport, all while building upon the IDPF standard's foundation. The NetDev 0x18 Conference is set to be a significant event, showcasing advancements that will shape the AI network landscape. For more information on the conference, including abstracts for these and other offerings, please visit the NetDev 0x18 Conference website: https://netdevconf.info/0x18/index.html.