DPDK Summit Bangalore 2019

We, NTT Communications, have been working on research and development of a (hopefully) full-fledged software router called Kamuee. We will introduce how we integrate Poptrie, urcu, DPDK, CLI, frr, and others into Kamuee to provide an easy-to-use software router. We also share our experiences (difficulties) in using it in some large/famous events.

DPDK brings its value to add with performant data plane functionality. Kubernetes focuses on service velocity, simplicity, and scalability. One aspect of that simplicity, container networking by default has only one network interface – that too for control plane and management plane purpose.

DPDK plug-ins are vital in filling the gap in container Network Interface, CNI. Thus it enhances the performance metrics of container networking with these DPDK plug-ins.

Through this presentation, attendees like ecosystem partners, Linux Open source developers will get an understanding of additional benefits of DPDK plug-ins such as shown below.

When using DPDK Virtual Machine with, say, open stack provides control at the granularity of Virtual machine but not at the multiple VNFs that may be running in VM.

Whereas with DPDK plug-ins, the granularity of control can be exercised at VNF level – with 100s or 1000s of VNFs, this granularity of control is very beneficial to VNF ecosystem partners.

Along with container’s scale-in and scale-out, DPDK plug-ins provide Enhanced Platform Awareness through plugins for
1) DPDK plug-in for Node Feature Discovery – for the ability to request and allocate resources
2) DPDK plug-in for CPU Kubernetes Manager - Support for CPU core pinning for K8 pods
3) DPDK plug-in for DPDK VHost-User – Support for Service Chaining aka East-West traffic -
4) DPDK plug-in for SR-IOV – Support for Accelerated packet processing -

The ecosystem partners and developers will take home the GitHub resources to follow through with their development. To give one sample, Kubernetes Vhost user CNI
https://www.google.com/search?q=Kubernetes+VHOST+CNI&rlz=1C1GCEA_enUS773US773&oq=Kubernetes+VHOST+CNI&aqs=chrome..69i57.10549j0j7&sourceid=chrome&ie=UTF-8

And likewise for each of the DPDK plug-ins. 

User Plane Function(UPF) handling the data plane traffic of 5G is responsible to handle various types of traffic. The UPF is expected to be implemented in virtual form considering the horizontal scalability and in a cloud native way. One of the choice of implementation is using DPDK and FD.io components. GTP-U is realized using the VPP and QoS enforcement check for the uplink (UL) and DSCP marking for the downlink(DL) is realized the using the DPDK pipeline. The implementation is planned in such a way that it is horizontally scale considering the higher throughput and performance need. Considering the stringent performance numbers of 5G application UPF instances are expected to handle independently certain specific flows / services pertaining to a gNodeB or group of gNodeBs. In each of the UPF instances the flow based QoS mapping based on the 5QI (QFI) value of the GFP extension header used for classification, marking and rate limiting. The SR-IOV capability of the NIC interface is used to classify and segregate the traffic to the UPF instances. This concept also cover the number of hierarchies from the QoS aspects and the approach to map the DPDK H-QoS scheduler and also from the shaping perspective for the non GBR flows to gurantee the GBR flows.

In this talk, we will introduce FD.io VPP and the Ligato cloud-native networking framework and demonstrate how these are used to address two use cases - a CNI plugin for Kubernetes (Contiv/VPP) and an IPSEC VPN gateway.

FD.io VPP is a high-performance packet forwarder that runs on commodity CPUs. Ligato provides a platform for developing Cloud-Native Network Functions (CNFs) using VPP running in Linux user-space and leveraging Intel's DPDK to grab packets directly from the NIC.

Contiv/VPP is a CNI (Container Network Interface) plugin that uses VPP to provide network connectivity for Kubernetes pods. Its aim is to bypass the kernel for packet handling in a Kubernetes cluster wherever possible. To achieve this, it replaces the kube-proxy load-balancer by leveraging VPP's advanced NAT functionality, and implements native Kubernetes policy (to avoid the requirement for a kernel-based policy plugin). As well as providing fast packet processing, Contiv/VPP can be used as the foundation for CNF deployments running in Kubernetes - since it is based on the Ligato VPP Agent its functionality can be easily extended to interconnect CNFs, e.g. using the Ligato SFC Controller.

StrongSwan is probably the most common IPSEC VPN concentrator in use today. It provides an IKE daemon but leverages the Linux kernel for IPSEC forwarding. Using Ligato we can interface StrongSwan's IKE implementation to VPP's IPSEC forwarding plane to create a high-performance IPSEC VPN CNF.

In general we will look at the benefits of forwarding traffic using VPP rather than the Linux Kernel in terms of: 1) performance 2) flexibility 3) manageability

This presentation will give an overview of how a new OVS action can be added
with an example. It will cover
- how to implement a new action
- which parts of code need modifications.
- how to add this new action in kernel datapath and DPDK userdatapath
- an overview of how flows are matched and translated by Ovs-vSwitchd
     when an upcall happens.

Cover solutions/innovations that are evolving in both software and hardware architectures to support data plane for latest technology needs. As part of the discussion plan to cover challenges observed in designing data plane to support 5G features, cloud native environment and edge location.

DPDK based applications are modeled to work as 'stand alone or software library' in 'run to completion or pipelined' for 'single or multiple process models'. It is a tedious task to isolate, debug and understand odd behavior which occurs randomly or periodically.

The goal of the talk is to share some of the insights, effort in documenting these and changes done in procinfo tool to help the same.

This presentation will talk about optimized locks and new locks in a step-wise way:
1. Optimizing the generic spin locks and read-write locks to save the CPU cycles of contending threads.
2. Implement the arm64 specific wfe instrution based locks to save interconnect traffic and power.
3. Implement the new ticket lock to ensure the fairness by obtaining the locks in the requesting order.
4. Implement the MCS algorithm based queued spin lock to decrease cache snooping traffic and help performance.
Number #2 is aarch64 specific and the other three are generic and help all the architectures.

This presentation will go over the user level experience of building, configuring and running DPDK based iPerf - https://github.com/ansyun/dpdk-iperf


With DPDK based applications growing, it is paramount to have tools to measure the performance of DPDK based appliances / VNFs. Any traditional iperf /netperf tools are kernel based. So, using them does not measure DPDK based solutions since the kernel based tools peak out and tools become a bottleneck.

There are many DPDK based traffic generator available – most of them are used for stateless traffic generation. So, their documentation is light on stateful traffic generation and measurement. In addition, many have both client and server in the generation and measurement modules themselves. This offers the challenge to terminate and measure performance at the DPDK based servers – like DPDK based DPVS for instance.

iPerf is very familiar to all network developers. And now that iPerf is available on userspace, this presentation will go over the user level experience of building, configuring and running DPDK based iPerf - https://github.com/ansyun/dpdk-iperf

Express Datapath zero copy is one of the packet acceleration methodologies in kernel. As and when new vulnerabilities and security threats are reported, patches are been updated.

While running with user desired BPF/eBPF, XDP-ZC PMD is an attempt to allow DPDK applications to run with all the security benefits, less memory copy overhead and hardware offload capabilities.

As the world becomes ever more digitally and globally connected, industries are gearing for transformation through the adoption of disruptive digital technologies such as IoT, Artificial Intelligence (AI), Augmented Reality (AR), and Virtual Reality (VR). The success of this transformation is almost entirely dependent on the ability of networks to meet the diverse Quality of Service (QoS) demands, such as ultra-low latency, high throughput, or massive IoT. 5G is addressing these stringent business critical QoS requirements of various industry verticals. Network Slicing is a key capability of 5G that enables the flexibility to meet the differentiated network performance needs of these industries, by creating multiple logical or virtual networks or slices out of common physical network infrastructure. Network equipment providers, to implement network slicing, are leveraging the cloud-native architectural paradigm to provide agility, scalability and exploring DPDK, OVS or FD.io for providing network programmability to 5G networks to meet the QoS needs of various industry verticals.

TCS' 5G Network Slicing reference solution is built on its  Microservices Development and Deployment (MDD) framework leverages OVS enhanced to support GTP protocol for data plane implementation. The framework also provides a powerful zero-touch orchestration engine to ease complex multi-region deployments and life cycle management to the solution.