- people - netdevconf.info

0x15: Bronze sponsor, NVIDIA!
by Jamal Hadi Salim 01 Jul '21

01 Jul '21

We are pleased to welcome NVIDIA (https://www.nvidia.com/) as a bronze sponsor for Netdev conf 0x15! NVIDIA is the leader in open networking with end-to-end solutions for SmartNICs, DPUs, open optics, and accelerated data center switches. NVIDIA is committed to open-source at all layers of the software and hardware stack with open networking initiatives like Linux Networking, SwitchDev, DENT, and the SONiC network operating systems and believes that open-source has the power to accelerate innovation and benefits everyone. cheers, jamal

1 0

0x15: New Talk accepted!
by Jamal Hadi Salim 01 Jul '21

01 Jul '21

Tom Herbert is still on a quest to speed up the network stack[1]. In this talk he introduces the PANDA parser with intention to replace the venerable Linux kernel Flow Dissector. Tom says the flow dissector is hard coded, convoluted making it both hard to extend and hard to maintain. (yes, he is guilty as well having been one of the originators of the flow dissector). PANDA parser is a domain specific parser that lives under the philosophy of "write once, run anywhere, run well". Unlike Flow Dissector, a PANDA Parser with metadata extraction is written in a declarative representation as opposed to imperative instructions - all in familiar C. It has been shown that while more flexible, the PANDA parser is more performant than flowdissector. The PANDA parser may be compiled to different backends, currently two implemented backends are available: an optimized userspaces C and an XDP/eBPF one. There is ongoing work on generating a plain kernel version as well which may be consumed by other part of the kernel. For any of those 3 backends, the parser definition stays unchanged. [1]https://legacy.netdevconf.info/0x14/session.html?talk-BP4-byte-code-for-programmable-protocol-independent-PDU-processing More info: https://netdevconf.info/0x15/session.html?Replacing-Flow-Dissector-with-PAN… cheers, jamal PS: Registration is open, see: https://netdevconf.info/0x15/registration.html PPS: We are looking to post the schedule today

1 0

0x15: New Talk accepted!
by Jamal Hadi Salim 01 Jul '21

01 Jul '21

Marcelo Leitner and Vlad Buslov will discuss turbo-charging the TC flower control interface and the evolution of the work that went into it over a period of two years: At least one hundred kernel patches, many changes to user space and increase in performance by a magnitude. To the datapath skeptics who are wondering why optimize the control path (reminds of the joke from not-to-be-named kernel guy under the influence "Why do we need userspace at all?"): You do care if you are constantly churning or accessing the datapath. Example workloads: - short lived container workloads where you have to add/delete access control rules, - telco environments, where efficient rule statistics are critical billing - etc. cheers, jamal PS: Registration is open, see: https://netdevconf.info/0x15/registration.html PPS: We are looking to post the schedule today

1 0

0x15: New Talks accepted!
by Jamal Hadi Salim 30 Jun '21

30 Jun '21

1) Alex Aring asks if you are paying attention to socket ECONNRESET. Big consequences if you are trying to be reliable as Alex et al recently found with the kernel Distributed Lock Manager (DLM) protocol. In this talk he will make suggestions on changes that make this less painful. More info: https://netdevconf.info/0x15/session.html?ECONNRESET-on-a-reliable-socket-a… 2) sFlow is an established industry standard for real-time network monitoring which has seen wide adoption in switching ASICs. In this talk Peter Phaal, Neil McKee, Ido Schimmel, Roopa Prabhu, and Andy Roulin will delve into how sFlow can be used to monitor a data center components ranging from switches to VMs. They will discuss sFlow data formats and their recent extensions. More info: https://netdevconf.info/0x15/session.html?Network-wide-visibility-with-Linu… 3)Simon Horman will discuss the new enhancements to the TC police action to add packet based control. He will describe how the feature gets offloaded and how it gets to have its own lifecycle independent of the classifier. More info: https://netdevconf.info/0x15/session.html?Recent-Enhancements-to-the-TC-Pol… 4) New feature testing or bug reproduction can be cumbersome as one embarks on creating network environment scripts - defining network namespaces, setting up different network interfaces types, assigning addresses, configuring routes, vrfs, etc.. All this is tedious and error prone and often frustrating. Eyal Birger introduces netpen.io to will save your valuable time ane energy. More info: https://netdevconf.info/0x15/session.html?netpen.io-visual-editor-for-netwo… cheers, jamal

1 0

0x15: New Talk accepted!
by Jamal Hadi Salim 29 Jun '21

29 Jun '21

New talk accepted by the illustrious PC. Marta Plantykow, Milena Olech and Maciej Machnikowski are back again discussing network infrastructure performance tweaking using AI[1][2], this time with a focus on PTP protocol. Network cards that are equipped with PTP Hardware Clock (PHC) rely on the PTP protocol to synchronize. There is that additional step of synchronizing the system time to the PHC. Linux ptp applications use a feedback control system (PI controller) to even out timing errors. In this talk Marta et al use a genetic algorithm (GA), a stochastic algorithm used in the AI world, to optimize the parameters of the PI controller. They explore different approaches to solve the optimization, various data evaluation methods and results achieved. Come hear, and engage them. cheers, jamal PS: Registration is now open - see : https://netdevconf.info/0x15/registration.html PPS: We are working on the schedule, coming RSN.

1 0

0x15: Registration opened!
by Jamal Hadi Salim 28 Jun '21

28 Jun '21

It is our pleasure to announce the opening of registration for Netdevconf 0x15. Sorry it took us this long - we are trialing a new platform this time so we wanted to make sure all goes smoothly. The fee is USD $50. Students are 50% off. Note that when registering, your Name, Email, and Country fields are mandatory for you to fill in, designation and company name are not. Please use your proper name. For more details, go to: https://netdevconf.info/0x15/registration.html cheers, jamal

1 0

0x15: New Talks accepted!
by Jamal Hadi Salim 28 Jun '21

28 Jun '21

4 more talks accepted by the illustrious PC! 1) Nabil Bitar and Jamal Hadi Salim introduce Policy Tables (Ptables), an Access Control List (ACL) infrastructure inspired by the trio: IPtables, IPSet and Conntracking. The IPtables trio is very widely deployed and well understood from a management perspective but has ongoing performance challenges[1]. Ptables maintains the IPtables management semantics while overcoming the IPtables performance challenges. Ptables is implemented using eBPF at TC and XDP levels; however, it is intended to use whatever kernel mechanisms available to meet desired goals. Alternative proposals take an _evolutionary_ approach by keeping IPtables ABI compatibility (either by adding kernel or libiptc transform layers); they have so far failed to see wide adoption. The authors instead take a _revolutionary_ approach by ignoring the ABI altogether and focusing on the semantic equivalence presented to the IPtables crafts-person. Ptables does not strive to be a full replacement for IPtables; instead, it takes a minimalist approach to cover the most common use cases while leaving room for innovating on new ideas. In this talk the authors will discuss the architecture and design choices taken as well as implementation challenges faced fitting eBPF. They will also discuss future plans. 2)Ping and Traceroute have served us well over the years, however they have been overcome by events in the networking world and do have shortcomings. Ron Bonica discusses revamping these tools. Ron describes the improvements over ping in the new PROBE diagnostic tool described in IETF RFC 8335. He also discusses ICMP extensions detailed in IETF RFC 5837 to enhance traceroute. Ron will further discuss the contributions they have made to the Linux kernel and user space tooling to enable both these extensions. The talk will describe the tooling details and contributions involved. 3) Qizhe Cai, Shubham Chaudhary, Midhul Vuppalapati, Jaehyun Hwang and Rachit Agarwal have spent a year analyzing Linux network stack performance issues in a data centre environment. While a lot of prior performance evaluations are centred around short-lived TCP flows, studies show that data centre flows mostly tend to be lived. This changes the reality of where the bottlenecks are. Example, the authors found that in fact the dominant bottleneck is kernel<->user copies as opposed to TCP processing which reveals its head when the focus is on short flows. The study also looks at, amongst other things, the effect of ever increasing data centre wire speeds(and the resulting bandwidth-delay products) on modern CPU cache sizes and calls out for new mechanisms for network-aware CPU scheduling. The authors will make recommendations and discuss challenges met. 4) Maxim Mikityanskiy describes how to accelerate synproxy with XDP. SYN cookies are a popular stateless DOS mitigation mechanism in which the sequence number of the first ACK is chosen as the result of a cryptographic hash function over easily computed parameters. Though effective, in Linux this requires traversing the network stack all the way to the socket layer. The extra compute cycles limits the rate of issuing SYN cookies. Maxim describes using eBPF at the XDP level to check SYN cookies without binding to a socket. The effort includes adding new eBPF helpers to aid this effort. [1] https://netdevconf.info/0x15/session.html?talk-linux-ACL-performance-analys… cheers, jamal

1 0

0x15: New Talks accepted!
by Jamal Hadi Salim 27 Jun '21

27 Jun '21

4 more exciting talks accepted by the illustrious PC. 1) Ido Schimmel and Petr Machata will discuss the kernel 5.13 feature supporting resilient nexthop groups. Resilient NH groups are intended to minimize disruption in flow routing when changes are made to the group composition and weights of constituent nexthops. This talk will detail the inner workings of resilient nexthop groups, including hardware offload. The talk will also cover future work such as static resilient nexthop groups, Nexthop group consolidation, and integration into FRR. 2)Kostis Kaffes, Jack Tigar Humphries, David Mazières, and Christos Kozyrakis address the notion that OS schedulers often, if not always, have discrepancy in their algorithms that do not take care of what networked applications need because they lack system level insight into those needs. There is a lack of coherency in the kernel between the scheduling and the various layers, including the network stack, NICs, and applications. The result is suboptimal scheduling decisions which are often solely responsible for applications poor latency and throughput. Kostis et al propose "Maple" - a framework for user-defined scheduling. A user expresses an application-specific scheduling policy in a few lines of ebpf code and then deploys it across system layers without modifying their code. Kostis et al claim up 8× performance improvement compared with default policies. 3) Mark Claypool, Benjamin Peters, Pinhan Zhao, and Jae Won Chung evaluate TCP Hystart performance on satellite links. To avoid overshooting TCP slow start ramp up, the Linux kernel has a Hystart mode (on by default) that will exit slow start before packets are lost. Mark et al demonstrate that Hystart-Delay often causes spurious exit of slow-start on the satellite links they tested. This has big implications for efficient utilization of commercial satellite Internet links. Mark et al will be seeking suggestions from the community on how to overcome these challenges. For the moment, they have a suggested configuration which is being used as a short term solution. They will also briefly discuss future work to address the issue. 4) Eric Dumazet is not a man who uses many words. The title of his talk is simply "Big TCP". In this talk he discusses challenges faced when the network stack now has to deal with 200/400Gbit speeds. How to address this? Once one reduces or avoids use<->kernel copying the size of TSO/GRO packets matters. In this talk Eric will present how we can override current limits and reduce TCP/IP stacks overhead. cheers, jamal

1 0

0x15: New Talks accepted!
by Jamal Hadi Salim 26 Jun '21

26 Jun '21

4 new exciting talks have been accepted by the PC 1) Alex Markuze takes another crack at dealing with socket overhead. Most solutions overcome the overhead by moving the processing to user space via zero copy. Alex will describe a new approach called MAIO with kernel/userspace shared memory pages from a dedicated allocator. Unlike DPDK, the kernel tcp stack is used by means of dedicated rx/tx kernel threads+kernel tcp sockets. 2) Jesse Brandeburg and Brett Creeley describe PowerMAN, a user space tool for Dynamic Interface Power Management. NIC hardware buffering is not keeping up with increasing speeds. Achieving maximum performance implies tuning the NICs which typically comes at the cost of power consumption. PowerMAN addresses these issues. Jesse and Brett are looking for feedback: Should this tool move in-kernel or stay in user-space? Much of the work may be more appropriate for the kernel scheduler to be involved. 3) Boris Pismenny, Yoray Zack, Ben Ben-Ishay and Or Gerlitz will discuss work on Autonomous NVMe-TCP offload. Traditional approaches to offloading NVMe-TCP require offloading a lot of L4 functionality: TCP, IP, routing, QoS, NAT, firewall, etc Not only is this undesirable given the complexity - but is also easily mitigated for bulk storage using batching as demonstrated in 0x14 talk[1]. Boris et al present how to offload CPU-intensive operations that cannot be optimized away using batching or clever software engineering: copy and CRC independent of L4 functionality. 4) Nabil Bitar and Jamal Hadi Salim describe their performance work on Linux ACL subsytems. Their work was motivated by lack of existing extensive comparison that fairly evaluate both _forwarding_ and _host_ workloads. The comparison is modeled using a widely deployed iptables use case. They look at: Iptables(with and without conntracking), IPtables with IPSet (with and without conntracking), XDP/ebpf, tc/ebpf, and tc/flower. They analyse both control path as well as data path performance under a variety of conditions. They are hoping their experience will help others in the community make decisions as to what approach to take under given circumstances. They are hoping as well to get community feedback and evolve the work further. cheers, jamal [1] https://netdevconf.info/0x14/session.html?talk-NVMe-over-TCP-is-not-NVMe-ov…

1 0

0x15: New workshop and Tutorial accepted
by Jamal Hadi Salim 25 Jun '21

25 Jun '21

Two new sessions accepted. 1) Alexander Lobakin and Jesper Dangaard Brouer will chair the XDP workshop. Details: https://netdevconf.info/0x15/session.html?workshop-XDP-general 2) First tutorial on an extremely important but often over-looked topic in the Linux Network stack, Time Synchronization over ethernet. In this tutorial Maciej Machnikowski will go from introduction to details on the topic. Details: https://netdevconf.info/0x15/news.html?time-sync-over-ethernet-tutorial-acc… cheers, jamal

1 0

2024

2023

2022

2021

2020

2019

2018

people