Increase Application Performance with SmartNICs€¦ · Cloud networking using Open vSwitch,...

© 2016 NETRONOME SYSTEMS, INC.

Ron Renwick, Sr. Director of Product Management

Increase Application Performance with SmartNICs

© 2017 NETRONOME SYSTEMS, INC. 2

Comprehensive SDN and NFV Acceleration

Accelerates performance in NFV and Cloud infrastructures

Low profile PCIe Gen3x8 Adapters 10/25/40/100GbE solutions

Agilio® CX SmartNICs

§  Boosts network performance and frees up x86 CPU cores

§  Brings the speed of software innovation to hardware

§  Cloud networking using Open vSwitch, Contrail vRouter and eBPF applications

Agilio® Server-Based Networking Software


Delivering Network Accelerators for High Growth Markets

NFV Infrastructure Cloud Networking

Security


The Speed of Software Innovation on Hardware

Rapid availability of new networking features

New production-level feature rollouts at needed performance every 4-6 months

© 2017 NETRONOME SYSTEMS, INC. 5 5

Source: Ericsson

Cloud Adoption Driving New Server Workloads

76% Data center traffic driven by cloud-based workloads by 2018

Source: Cisco Global Cloud Index, 2015-2018

Diverse and Emerging Data Center Workloads

Networking Workloads Other Workloads

Distributed Storage (SSD, HCI)

Network Function Virtualization (NFV) Distributed Security Packet Processing Machine Learning

Web Server & Search Transation Processing Scientific Computing


Workloads that Benefit with Accelerator

Scientific Computing

Packet Processing

Machine Learning

Deep Learning

Distributed Security


GPU Accelerator

Distributed Storage

Machine Learning

Packet Processing

Workload optimized accelerators work as co-processors to general x86 compute

FPGA Accelerator

Network Accelerator

NFV


6


Open vSwitch Datapath Acceleration

OpenStack Neutron

OVS ML2

Compute Node

VM VM VM VM . . . OpenStack Nova Agent

OVSDB OpenFlow

Linux Kernel

Agilio CX OVS Datapath

Actions Match Tables

OpenStack Nova

Open Daylight Controller

(ODL)

Tunnels

Deliver to Host

Update Statistics

Seamless Integration with Control Plane

OpenStack Ready

Transparent Offload

SR-IOV & XVIO

Connectivity to VMs

ovs-vswitchd

ovs-dbserver

OVS Datapath

Actions Match Tables Tunnels

https://www.netronome.com/media/redactor_files/WP_Agilio_SW.pdf

P4,C Data Path Extension


Contrail Neutron Plugin

Integration with OpenStack and Contrail

VM . . . OpenStack Nova Agent

vRouter Agent (user space)

XMPP

Linux Kernel

vRouter Forwarding Plane (Kernel Module)

Netronome Agilio CX

MPLS Label Lookup

Outer LPM and Next Hop

Contrail Node

Decap Flow Table Policy

Deliver to Host Update Analytics

Configuration Node Mirantis

OpenStack NETRONOME

SR-IOV and Virtio

Server

Netronome SmartNIC

VM VM VM

OpenStack Nova

COMPUTE NODE

26


Agilio CX SmartNIC

Per Server CPU Core Efficiency with Agilio Platform

Virtualized Server with Tenant VMs and

Apps

Agilio CX SmartNIC

50X Efficiency Gain vs. Kernel OVS

20X Efficiency Gain vs. User OVS

Open vSwitch (OVS)

Full OVS Offload

SDN Controller


NFV Use Case: 2,000Kpps per VNF or Application

Rack Throughput: 168Mpps VNFs Per Rack: 80

Server TOR

Server Server Server Server Server

Server Server Server Server Server Server Server Server Server

20 S

erve

rs w

ith 2

x40G

bE

Server TOR



20 S

erve

rs w

ith 2

x40G

bE

Rack Throughput: 440Mpps VNFs Per Rack: 220

Racks Needed to Support 220 VNFs Racks Needed to Support 220 VNFs

Server TOR



Server TOR



Server TOR



Server TOR



2.8

C C C

C C C

C C C

C C C

C C C

C C C

C C C

C C C

OVS 16 Cores

9.6 Mpps of VXLAN

Processing

4 Apps or VNFs at

2000Kpps

C C C

C C C

C C C

C C C

C C C

C C C

C C C

C C C OVS

VMs 23 Cores

22 Mpps of VXLAN

Processing

11 Apps or VNFs at

2000Kpps

VMs 8 Cores

Server Core Allocation Server Core Allocation

Ericsson Cloud SDN with OVS in User Space and Traditional NIC

Ericsson Cloud SDN with OVS running on Netronome Agilio Platform

3X Lower TCO


P4 and eBPF-based Rapid SDN Innovation


Accurate, High Performance Network Telemetry

With Agilio Solution

Real time data collection

In packet, wire-speed

Physical & virtual network

Triangulate server, VM, switch

Gather data at each hop

Increase Service Levels with Inband Network Telemetry


Comprehensive SDN and NFV Acceleration

Delivers compelling TCO in NFV and Cloud infrastructures

Low profile PCIe Gen3x8 Adapters 10/25/40/100GbE solutions

Agilio® CX SmartNICs

§  Boosts network performance while freeing up x86 CPU cores

§  Brings the speed of software innovation to hardware

§  Cloud networking using Open vSwitch, Contrail vRouter and eBPF applications

Agilio® Server-Based Networking Software




Subject

Using OpenStack to increase application performance with SmartNICs

SmartNICs improve performance by offloading and accelerated the datapath for server based networking. We will discuss controlling and managing SmartNICs in an OpenStack environment to provide increased application and network performance.


Sources: Intel, Altera, AMD

GPU Accelerator

FPGA Accelerator

Network Accelerator

Workloads TAM Shifting from x86 to Accelerators


Accelerators for Efficient Heterogeneous Computing


Packet Processing

Machine Learning

Deep Learning



GPU Accelerator

Distributed Storage

Machine Learning

Packet Processing

Workload optimized accelerators as co-processors to general x86 compute

FPGA Accelerator

Network Accelerator

NFV



Custom Datapath with P4/C Integrated Dev Env (IDE)

Seamless programming of SDN applications into the production Agilio solution

Utilizes open source P4 compiler, language from the P4 Language Consortium

Extensions enable optional C-based modules for sophisticated functions

P4/C IDE is generally available. Industry-wide open R&D at www.Open-NFP.org

Agilio Intelligent Server Adapters 2x10GbE , 2x25GbE, 2x40GbE

P4 Code

C Code (optional)

Netronome Programmer Studio

First in the Industry – on production networking hardware


OVS Benchmark Throughput Results

OVS L2 Forward to VMs

Packet Size

Mill

ion

Pac

ket P

er S

econ

d

OVS VXLAN + L2 to VMs

Packet Size

Mill

ion

Pac

ket P

er S

econ

d

•  1000 Wildcard Flow Rules (match on MAC DA) •  64,000 Exact Match Micro-flows •  Bursting with 10 packets/Micro-flow

•  1000 Wildcard Flow Rules (Match on VNI+MAC DA) •  64,000 Exact Match Micro-flows •  Bursting with 10 packets/Micro-flow

https://www.netronome.com/media/redactor_files/WP_OVS_Benchmarking.pdf

With one x86 CPU Core

With one CPU Core

With 10+ x86 CPU Cores

With 10+ x86 CPU Cores


Custom Features Dev Entails Unpredictable ROI

SDN and NFV require many networking features

Data center operators and OEMs want to differentiate

Develop on x86 using C, then port to programmable networking hardware

•  Unpredictable or poor performance

•  Long time to market

•  Vendor lock-in HDL on FPGAs

C code on x86

C code on MIPS or ARM SOCs

Micro code on NPUs

17


Programmable Networks A New Era

New algorithms and application architectures demand specialized compute

Data in-flight can be per-processed for highly parallel computations

Programmable NICs providing CPU offloads are cost effective method for increasing server compute capacity

Meets needs of IoT and NFV market

30


Accelerators for Heterogeneous Computing


Packet Processing

Machine Learning

Distributed Storage


GPU Accelerator

Distributed Storage

Machine Learning


Compression NVMe over Fabric CEPH

Firewall, ACL Crypto

Crypto

Workload optimized accelerators as co-processors to general x86 compute

SDN, NFV

FPGA Accelerator

Network Accelerator


From Free CPU Cycles to Application Starvation

Netronome solutions return expensive CPU cores to revenue generating VMs

CPU Cycles Used up for Networking With increasing adoption of SDN and NFV

CPU Cycles Available for Revenue Generating Applications

Per

cent

of C

PU

C

ycle

s in

a S

erve

r

5%

100%

80%

1990s 2000s 2010s 2020s

Cost per cycle increasing with slow down or demise of Moore’s law

Output per server dwindling as VMs and Apps are

starved for CPU cycles and I/O

bandwidth


Per Server CPU Core Efficiency

Throughput with single Server CPU Core

Mill

ion

Pac

ket P

er S

econ

d

•  50X Efficiency Gain vs. Kernel OVS •  20X Efficiency Gain vs. User OVS

https://www.netronome.com/media/redactor_files/WP_OVS_Benchmarking.pdf


Large R&D budgets, deep acceleration software expertise Proprietary silicon & hardware-based acceleration

Need for Accelerators Validated by Mega-Scale Operators

Everyone else needs off-the-shelf solutions


Custom Datapath with P4/C Integrated Dev Env (IDE)

Seamless programming of SDN applications into the production Agilio solution

Utilizes open source P4 compiler, language from the P4 Language Consortium

Extensions enable optional C-based modules for sophisticated functions

Agilio SmartNICs 2x10GbE , 2x25GbE, 2x40GbE

P4 Code

C Code (optional)

Netronome Programmer Studio

27

P4/C IDE is generally available. Industry-wide open R&D at www.Open-NFP.org

Date post:	09-Sep-2018
Category:	Documents
Upload:	phamtuyen
View:	219 times
Download:	0 times

Increase Application Performance with SmartNICs€¦ · Cloud networking using Open vSwitch,...

Documents