February 2025 - people - netdevconf.info

0x19: Talk, Reservoir to the tap: with fairness and humility
by Jamal Hadi Salim 09 Feb '25

09 Feb '25

Fintech workloads and few other modern work loads require very fine grained control (down to a couple of milliseconds) over packet pacing and at the same time be able to do this for millions of flows; all requiring different granularity of flow pacing. In this talk, Anjali S Jain, Vadivel Kannappa, Joshua Hay, Robert Hathway, Magnus Karlsson, Chihjen Chang, Naren Mididaddi, Priyalee Kushwaha and Milena Olech go into implementation details of how to achieve these requirements. Details: https://netdevconf.info/0x19/32 cheers, jamal

1 0

Communication via Internal Shared Memory (ISM) - Time to open up
by Jamal Hadi Salim 09 Feb '25

09 Feb '25

Internal Shared Memory (ISM) is a method for synchronous, access controlled, direct memory access into dedicated memory regions between systems that share the same physical memory. With increased popularity of using shared memory semantics over the network protocols (see all the AI/ML craze going on), this talk will shed some light on these many years of experience on huge mainframes and other related efforts. In this talk Alexandra Winter will demystify for us ISM, its features and usecases, and explain several proposals that have been made to extend the usage of ISM. Alexandra will further discuss what would be the best way to provide clean interfaces in the Linux kernel, to enable future extensions. An early RFC has been sent to the netdev mailing list with a proposal for an ism layer - to finally make it open to the masses. Details: https://netdevconf.info/0x19/12 cheers, jamal

1 0

0x19: Talk, Optimizing Bandwidth Allocation with ETS and Traffic Classes
by Jamal Hadi Salim 09 Feb '25

09 Feb '25

In the current deployed approach in Linux, the operator does not have much control over how wire bandwidth is allocated between VFs. For example, when provisioned, a 50Gbps PF's equally split between two VFs, giving each 25Gbps, with traffic classes (e.g., TC0 and TC1) limited to percentages of these smaller allocations. Carolina Jubran says implementing ETS hardware support allows for control over such allocations. Using TC bandwidth allocation, the PF can instead be divided directly into 80% for TC0 and 20% for TC1 across the two VFs. In this talk, Carolina will describe the insufficiency of existing approaches to solve these problems, reference previous work, and demonstrate how this implementation improves resource allocation across VFs by applying traffic class-based bandwidth proportions, supported by real-world results and a demo. Details: https://netdevconf.info/0x19/14 cheers, jamal

1 0

0x19: Talk, mq-cake: Scaling software rate limiting across CPU cores
by Jamal Hadi Salim 09 Feb '25

09 Feb '25

Qdiscs rely on global lock to sync state across CPUs and therefore dont scale in presence of many cores (or in presence of very high bandwidth). Jonas Köppeler, Toke Høiland-Jørgensen, and Stefan Schmid implemented a multi-queue variant of sch_cake that can scale its rate limiting across hardware queues (and thus CPU cores) by sharing a bit of state on top of the mq qdisc. In this talk, they will present the implementation, performance evaluation, as well as discuss their proposal for an API that will make this work upstreamable, and applicable to other qdiscs as well. Details: https://netdevconf.info/0x19/16 cheers, jamal

1 0

0x19: Talk, Linux Kernel Support for IOAM Direct Exporting
by Jamal Hadi Salim 08 Feb '25

08 Feb '25

In Situ Operations, Administration, and Maintenance (IOAM) provides a way to collect telemetry data on devices, e.g., switches and routers, in so-called limited domains (e.g., a datacenter or a service provider). Recently standardized in the IETF, IOAM has been developed with various operation modes (i.e., option-types) in order to be suitable for as many potential use cases as possible. In this talk, Maxime Goffart, Justin Iurman, Emilien Wansart and Benoit Donnet describe the IOAM Direct EXporting (DEX) option kernel implementation which allows each network device on the path to locally aggregate and/or export telemetry data towards one or more collector(s). The implementation comes with support in ubiquitous user space tools. The authors will the implementation details and challenges as well as present performance results. Details: https://netdevconf.info/0x19/11 cheers, jamal

1 0

0x19: Talk, Open Source Tooling for PHY Management and Testing
by Jamal Hadi Salim 08 Feb '25

08 Feb '25

One of the techniques the ethernet physical layer uses to achieve high speed rates is pulse-amplitude modulation(PAM). An implementation of a specific physical layer is commonly referred to as PHY. If you want to adjust the speed on the wire from say 1Gbps to 800Gbps (caveat emptor!) you pick a different set of signal ranges to be used for equalization for different combinations of bits on the wire. When a vendor ships you a NIC, these ranges will be embedded in the firmware and are fixed. If you have to interop then the idea is the peer NIC vendor has similar settings. How do you test these ranges? Lee Tragger says there are no open source tools available. Essentially, vendors provide their own proprietary tools to set equalization values and run pattern tests. In this talk, Lee will discuss whether setting equalization values and enabling pattern tests has matured enough across various vendors to develop a standard open source interface. Details: https://netdevconf.info/0x19/37 cheers, jamal

1 0

0x19: Talk, Democratizing Far Edge Networking: Journey of Mainlining Wired and Wireless in Linux Kernel
by Jamal Hadi Salim 08 Feb '25

08 Feb '25

In this talk, Robert Marko will discuss real-world experiences and lessons learned from mainlining support for various wireless and wired SoCs, including Qualcomm's IPQ40xx platform, Marvell's Prestera switches, Microchip Sparx-5 and LAN969x family of products, as well as upcoming work on MaxLinear's SoCs. Robert will explore the challenges and benefits of using upstream Linux for edge networking, including the evolution of frameworks like switchdev for network switch ASICs and the wireless subsystem for access points. The talk will cover specific examples of upstreamed patches, discussing the technical aspects of bringing up new hardware support while maintaining clean abstractions. The speaker will examine how standardized Linux interfaces like switchdev can replace traditionally proprietary stack while providing equivalent functionality. Details: https://netdevconf.info/0x19/35 cheers, jamal

1 0

0x19: Talk, The Battle Of The ZCs: Who is the prettiest of them all?
by Jamal Hadi Salim 07 Feb '25

07 Feb '25

In this talk, Nabil Bitar, Jamal Hadi Salim, Victor Nogueira and Pedro Tammela investigate the performance characteristics of two new high-performance application networking mechanisms in Linux: TCP devmem and IOURING zero copy. While TCP devmem is designed for zero copy to/from network/accelerators (such as GPUs), in order to compare to IOURING, the authors look at it from the perspective of client-server application; in the case of devmem, copying to/from user memory (udmabuf). Both these techniques are new to the Linux world and are still not fully integrated into Linux proper at time of the study - making them both challenging to deal with (kernel bugs popping up for example) and a bit of a mystery to most application writers. In addition to sharing the comparative results, the authors will shed some light for application writers to use these features. cheers, jamal

1 0

0x19: Talk, The future of SO_TIMESTAMPING
by Jamal Hadi Salim 07 Feb '25

07 Feb '25

SO_TIMESTAMPING is key to debugging the network stack and end-to-end app latency. SO_TIMESTAMPING is key to sifting bugs that are network vs those that can be attributed to the kernel. Jason Xing believes there are a few areas that need optimization for usability and performance. These include: uAPI compatibility, extra system call overhead, and the need for application modification. His work is to resolve these issues. In this talk, Jason will discuss his eBPF extension which transparently equips every matched flow with SO_TIMESTAMPING without _modifying any application code_. More details: https://netdevconf.info/0x19/5 cheers, jamal

1 0

0x19: Talk, IRQ Suspension: a new, efficient mechanism for packet delivery
by Jamal Hadi Salim 07 Feb '25

07 Feb '25

Traditionally apps have two choices: 1) trade increased CPU usage to get lower latency or 2) lower CPU usage at increased latency. The available choices were onerous because these mechanisms were only available interface-wide - that is until this work from Joe. In this talk, Joe Damato will discuss the IRQ Suspension feature which was recently merged in 6.13. The IRQ Suspension feature adds per-NAPI storage instances which enables more fine-grained tuning to the provided choices. Joe will describe the challenges overcame and explain when an application may decide to use the feature and how, exactly, to use it. Further discussion on current on-going work and possibly world production measurements, if available at time of the talk, will be shared. More details: https://netdevconf.info/0x19/4 cheers, jamal

1 0