- people - netdevconf.info

0x1A: Two talks on offloading security to DPUs
by Jamal Hadi Salim 22 Jun '26

22 Jun '26

In this talk Balakrishna Bhamidipati and Vijay Ram Inavolu describe total offload of TLS from the host onto a DPU. They run a reverse-proxy on the NIC to handle MCP traffic. This moves TLS termination, OAuth2/JWT validation, and session-aware L7 routing off the host and onto the NIC. The authors will describe the packet path, kTLS activation and fallback, session lifecycle, backend affinity, and SSE relay, showing how the design can be reproduced on commodity DPUs without any kernel modifications. https://netdevconf.info/0x1A/sessions/talk/macsec-protected-rdma-on-dpus-fr… In the second talk Alkama Hasan and Vijay Ram Inavolu describe how they resolve the hard problem of security for zero-copy mechanisms like RDMA. In such approaches, the payload bypasses the network stack altogether leaving the app outside the security fence that typically protects ordinary host networking. Large-scale AI and HPC jobs depend on RDMA to move data between nodes. MACsec is a natural fit to resolve this security challenge at layer 2. So what does such a DPU approach offer above the NICs that support MACsec offload today? Well, merely offloading MACsec lacks the ability to interact with a multi-tenant cluster orchestrator which a DPU-based approach can. IOW, an orchestrator (Kubernetes in this case) has no visibility into which nodes actually offer MACsec-protected RDMA egress, so a workload can't request placement on such a node. The authors illustrate this making MACsec become a schedulable property exposed in the node ResourceSlice via DRANet-Sec allowing workloads land on the right nodes automatically. https://netdevconf.info/0x1A/sessions/talk/macsec-protected-rdma-on-dpus-fr… cheers, jamal

1 0

0x1A: Talk, Can Homa and TCP Get Along?
by Jamal Hadi Salim 20 Jun '26

20 Jun '26

John Ousterhout examines TCP and Homa coexistence. Initial measurements show that when running TCP and Homa concurrently, TCP performs slightly better than Homa because Homa reduces its buffer utilization. However, over time Homa degrades badly because TCP selfishly fails to reciprocate and overloads buffers. John introduces homa_qdisc, a queuing discipline that paces both protocols' traffic, implements SRPT for Homa and limited SRPT for TCP, and balances output during congestion. Results? with homa_qdisc, both protocols improve. When running concurrently—Homa suffers only slight degradation, and TCP latency improves. Even better: TCP on its own rips the benefits of the homa_qdisc: with tail latency for short messages nearly halving versus fq_codel. Come and engage with John! cheers, jamal

1 0

0x1A: Talk, AF_XDP copy mode needs more love
by Jamal Hadi Salim 19 Jun '26

19 Jun '26

In the spirit of the legend Bob Marley (playing while i am sipping coffee and typing this!): AF_XDP: Could you be loved? Jason Xing is showing a lot of love to you. Most drivers can't do zero-copy, so copy mode is the universal path. Jason will talk about recent correctness fixes (multi-buffer TX leaks, continuation descriptors, TOCTOU metadata) cleared the bugs, while fine-grained profiling unlocked close to 2× throughput. Questions for you AF_XDP (yes, you!): Could you be locked - with finer grain? Could you be batched - without the pain? No single silver bullet! Just cycle-level perf, contention tracing, cache-line heatmaps. And the optimization love for you: - Send without budget blues - Hot structure reshuffles (cache-line dance) - Batch xmit — one ring to rule them - Doorbell coalescing (don't ring it twice) - Skb allocation acceleration - IRQ pairs trimmed short Don't let them fool ya — spinlock's not the ruler. Methodical, mon, methodical, one cache line at a time, thats how we get there! Ok, enough of my spin, read more details on the talk here: https://netdevconf.info/0x1A/sessions/talk/af_xdp-copy-mode-needs-more-love… cheers, jamal

1 0

0x1A: Talk, Networking Headless CXL Devices for AI Memory Services
by Jamal Hadi Salim 19 Jun '26

19 Jun '26

Another talk on the intersection of AI and memory requirements and how networking is a fundamental piece of the equation. AI workloads needs large, high-bandwidth, low-latency memory. Composable CXL memory fabrics address this by moving compute to data, running services such as search, caching, and quantization on smart Type-2 devices with local DDR. The trade-off? Device-side side needs IP endpoints, while CXL devices expose a memory window, not a NIC. In this talk, Vijay Inavolu and Gaurav Agarwal present a Linux virtual-interface path built from stock kernel pieces. Host and device daemons open /dev/net/tun, configure a virtual L3 interface, and mmap a shared CXL HDM-H window as a packet ring, enabling local apps to access memory with no kernel changes. The shared-ring design, CXL ordering rules, and a new host-pod bridge pattern give device-side Linux a cluster-facing service identity, so services are discovered and reached by service IP while clients stay unaware of the memory window. Measurements show 60x less host-link traffic than host-side compute and 2.56x FAISS throughput scaling across four cards. Vijay and Gaurav would like to find partners in this crime to build composable AI memory fabrics on top of Linux networking. Come, listen, provide feedback and yes collobarate! cheers, jamal

1 0

0x1A: Talk, The CXL Fabric End-Game: Bandwidth Realities and Networked Memory for AI Scale
by Jamal Hadi Salim 19 Jun '26

19 Jun '26

AI/ML workloads have a memory problem. The model memory requirements are growing exponentially but the GPUs just cant keep up. No worries; CXL makes gobbles of memory available. The RAM can be disaggregated and placed in racks outside the system reachable via a network link. PJ says CXL.mem is typically pitched as a memory expansion technology (treating it like a "far" NUMA node), but this framing understates a critical issue: CXL link bandwidth can be orders of magnitude lower than native DDR, not just higher latency. As AI workloads push toward distributed memory pools, this bandwidth gap fundamentally changes the cost model for scaling large models across disaggregated memory. PJ has some suggestions for upcoming CXL 2.0 to move beyond simple expansion toward true memory-coherent cluster execution, calling for coordinated work in switching infrastructure, firmware, and Linux kernel networking/memory subsystems. Come, listen and provide feedback to PJ! cheers, jamal

1 0

0x1A: Talk, Linux QUIC: Bringing a Modern Secure Transport into the Kernel
by Jamal Hadi Salim 18 Jun '26

18 Jun '26

Xin Long proposes migrating QUIC from user space into the kernel. He introduces QUIC as a native transport via a new IPPROTO_QUIC socket type enabling direct use by subsystems such as SMB and NFS. The architecture handles transport logic in the kernel while offloading TLS handshakes to user space, supporting POSIX-style APIs. The implementation has been validated through large interoperability tests and benchmarking. Come learn about this shiny new subsystem! More details: https://netdevconf.info/0x1A/sessions/talk/linux-quic-bringing-a-modern-sec… cheers, jamal PS:- Today is the last day for early bird registration i.e 50% off!

1 0

0x1A: Talk, TCP State of the union
by Jamal Hadi Salim 18 Jun '26

18 Jun '26

Eric Dumazet, aka "network ninja" will give the classical 2026 TCP State Of The Union. Recent and upcoming TCP changes to address performance scaling on modern platforms will be part of the highlights. Dont miss it! cheers, jamal

1 0

0x1A: BoF, Next generation Intel Ethernet Drivers
by Jamal Hadi Salim 17 Jun '26

17 Jun '26

Sridhat and JP will give us the innards of the nextgen Intel Ethernet drivers. ixd and idpf are Intel Ethernet drivers built around a common networking IP, spanning multiple product generations and form factors — discrete PCIe NICs, SoC NICs, SoC switches, and IPUs. idpf is an open-specification, unified PF/VF driver implementing the IDPF PCIe programming interface. It has been upstream in the Linux kernel for several years, is shipped in x86/ARM distributions, and is widely deployed in Google Cloud. ixd is a newer driver that combines control and data path functionality and leverages the switchdev framework to provision, configure, and manage devices that expose idpf-based data path functions. The BoF will cover the architecture and design of both drivers, the challenges of supporting diverse device modes, and the strategy for refactoring idpf to share code between ixd and idpf. It will also explain the different PCI function types (PF/VF) and subfunctions exposed by the hardware, and how the drivers are loaded to support the various possible configurations. Come and participate! More details: https://netdevconf.info/0x1A/sessions/bof/next-generation-intel-ethernet-dr… cheers, jamal

1 0

0x1A: BoF, Network Observability
by Jamal Hadi Salim 17 Jun '26

17 Jun '26

Jason Xing will lead a BoF on the exciting topic of network observability. Network observability has evolved from classic tools like `tcpdump` toward richer, layered capabilities that let engineers trace root causes after problems arise. The scope of this BoF is to discuss new ideas and survey existing techniques (which collect wide-ranging data points such as latency, throughput, skb drops, protocol diagnostics, reference counting, and many others). Specific sub-topics on the current agenda: 1. BPF timestamping 2.0: finer-grained, very low-overhead timestamping for broader use cases. 2. IETF TCP reset diagnostic payload: replacing current reset reasons with common APIs and messaging. Very welcom topics for discussion include how AI's rapid evolution may reshape network observability. Please ping Jason if you want to participate! More details: https://netdevconf.info/0x1A/sessions/bof/network-observability-bof.html cheers, jamal

1 0

0x1A: Talk, Resilient AI Supercomputer Networking using MRC and SRv6
by Jamal Hadi Salim 16 Jun '26

16 Jun '26

At very large scale AI training, tail latency and not average latency, is what limits synchronous pretraining performance. Christoph Paasch will discuss how MRC resolves these issues by: 1. Introducing a new RDMA transport protocol that sprays traffic across many paths and actively load-balances, eliminating flow collisions. 2. Creating Multi-plane Clos topologies and enabling >100K-GPU training clusters as two-tier networks with high switch radix and physical redundancy. 3. And last, using static SRv6 source-routing which gives MRC autonomous, failure-bypassing path control. The talk will describe experiences running MRC and static SRv6 routing in production within OpenAI and Microsoft’s largest training clusters where it has been used to train the latest frontier models. Christoph will discuss operational experience showing how MRC allows AI training jobs to ride out many network failures that previously would have interrupted training. cheers, jamal

1 0

2026

2025

2024

2023

2022

2021

2020

2019

2018

people