challenges in implementing high-performance sdn and nfv … · 2019-12-12 · sandeep shah,...
TRANSCRIPT
Sandeep Shah, Director of Systems Architecture, EZchip Technologies
Linley Data Center Conference, February 5-6, 2014
Challenges in Implementing High-Performance SDN and NFV Systems
© 2016 Mellanox Technologies 2
EZchip Overview
Fabless semiconductor company, NASDAQ listed (EZCH) Leading provider of Network Processors (NPUs) to the Carrier Ethernet (CE) market,
especially for edge routers EZchip is a strategic supplier to the top Carrier Ethernet vendors Recently announced the NPS, a revolutionary line of NPUs
• For next generation L2-7 router line cards, data center and cloud • Enabling new scalable architectures for SDN & NFV • 256 C-programmable Task Optimized Processors, 4K virtual threads • Large set of hardware and algorithmic accelerations • Integrates EZchip’s highly differentiated TM technology
EZchip founded in 1999; 200 employees, 160 in R&D in Israel Global offices in Israel (HQ); San Jose, CA; Boston, MA; and China Strong financial model; $191M in cash, no debt
© 2016 Mellanox Technologies 3
The Promise of SDN & NFV
SDN &
NFV
Improve network agility and operational efficiency
Reduce CAPEX & OPEX
Deliver secure functionality to thousands of tenants across dozens of datacenters
© 2016 Mellanox Technologies 4
New Network Architectures
SDN Decoupling data from control plane
NFV Decoupling software from hardware
Separate but complementary architectures
Data Plane Virtual networks
Control Plane Centralized
Software Virtual functions
Hardware Common platform
Virtualize also the network resources (in addition to compute and storage)
© 2016 Mellanox Technologies 5
Carrier Network Example
CPE
Customer Premise Equipment
CPE
CPE
CPE
PE
Provider Edge
PE
IP Backbone
Mobile Core
Network Functions
Network Functions
Network Functions
Enterprise Data
Center
Enterprise Branch
Enterprise Branch
Residences
security, load balance, DPI, monitoring
Build out data center capacity
Internet
Data Center
TOR Servers
SAN Spine Switches
© 2016 Mellanox Technologies 6
CPE
Customer Premise Equipment
CPE
CPE
CPE
PE
Provider Edge
PE
IP Backbone
Mobile Core
Network Functions
Network Functions
Network Functions
Enterprise Data
Center
Enterprise Branch
Enterprise Branch
Residences
security, load balance, DPI, monitoring
Data Center
Internet
Virtualizing mobile core functions
TOR
SAN Spine Switches
Carrier Network Example
Servers
vMobile Core
© 2016 Mellanox Technologies 7
CPE
Customer Premise Equipment
CPE
CPE
CPE
PE
Provider Edge
PE
IP Backbone
Network Functions
Network Functions
Network Functions
Enterprise Data
Center
Enterprise Branch
Enterprise Branch
Residences
security, load balance, DPI, monitoring
Internet
Virtualizing the network functions
TOR
SAN Spine Switches
Data Center
Carrier Network Example
Servers
vMobile Core vNFs
© 2016 Mellanox Technologies 8
CPE
Customer Premise Equipment
CPE
CPE
CPE
PE
Provider Edge
PE
IP Backbone
Enterprise Data
Center
Enterprise Branch
Enterprise Branch
Residences
Internet
vCPE Virtualizing the CPE
TOR
SAN Spine Switches
Data Center
Carrier Network Example
Servers
vMobile Core vNFs
© 2016 Mellanox Technologies 9
PE
Provider Edge
PE
IP Backbone
Enterprise Data
Center
Enterprise Branch
Enterprise Branch
Residences
Internet
Iaas / Paas
Data Center
TOR
SAN Spine Switches
Data Center
vCPE Virtualizing Enterprise DC
vNFs
Carrier Network Example
Servers
vMobile Core
© 2016 Mellanox Technologies 10
Implementation Challenges
NFV creating large server overhead • Most or all server cycles spent on networking functions • Performance does not scale much beyond 10G • Performance cliffs due to cache and TLB misses causing unpredictable behavior • vSwitch per server adds huge network management overhead
Fixed function networking silicon lacks critical capabilities • Lacks support for large number of addresses, flows, tunnels, buffering, security, QoS, sophisticated
traffic management • ACLs becoming more complex, not just basic 5-tuple anymore
- Requirements of 1K ACLs per VM, 40 VMs per server, 40 servers per rack yields 1.6M ACLs per rack - No switch can support that
• Unable to support new standards, features, protocols; features are burned in the silicon
© 2016 Mellanox Technologies 11
SDN & NFV Drive SW Solutions
Software solutions: Programmable
(e.g. CPUs & NPUs)
Hardcoded solutions: Fixed function ASICs
SDN & NFV
© 2016 Mellanox Technologies 12
SDN & NFV Network Challenges
Performance and Scalability: Data Path
Acceleration Load
Balancing Performance Monitoring
Many vSwitches
Many Subscribers
Reliability: High Availability
Faults & Correlation
Health Monitoring
Security: Tenants Services Functions
© 2016 Mellanox Technologies 13
Need to Accelerate
L2/3 switching & routing
Network overlay
termination
TCP termination
Security
DPI
Service chaining
Scaling to millions of
flows
Classification & ACL
Traffic Management
© 2016 Mellanox Technologies 14
Required
L2/3 switching & routing
Network overlay
termination
TCP termination
Security
DPI
Service chaining
Scaling to millions of
flows
Traffic Management
High-performance, feature-rich,
easily-programmed NPU to integrate within the
NFV & SDN framework
Classification & ACL
© 2016 Mellanox Technologies 15
NPS NPU for Smart Networks
L2/3 switching & routing
Network overlay
termination
TCP termination
Security
DPI
Service chaining
Scaling to millions of
flows
Traffic Management
Introducing the NPS
Classification & ACL
© 2016 Mellanox Technologies 16
NPS NPU performance
with CPU programmability
Introducing the NPS-400
400 Gbps C-programmable
Security and DPI hardware
acceleration Linux OS
Layer 2 - Layer 7 processing
© 2016 Mellanox Technologies 17
NPS-400 Key Features 400Gbps all layer C-programmable NPU
• 600Mpps wire speed with up to 960Gbps oversubscription On chip Traffic Manager & VOQ
• 1M queues, 5-level H-QoS 960Gbps of network I/O
• Including 1GE, 10GE, 40GE, 100GE, 400G ILKN and PCIe Gen 3.0 Integrated EFA (Ethernet Fabric Adaptor)
• Enables a full line card on a chip 256xCTOP processors with 4K HW threads at 1GHz core speed
• Native algorithmic instructions for efficient execution • 4K-way SMP Data Plane Linux, run to complete architecture with no SW overhead
Supports various inline services via dedicated HW accelerators • Security: IPSec/SSL encryption & decryption at 200Gbps • DPI: C-programmable RegEx stream content processing at 200Gbps
On-chip TCAM with TCAM algorithmic extension to external DRAM • Scales to millions of ACL rules
Based on commodity DDR3 & DDR4 memory providing up to 96GB • Virtually unlimited tables, states, counters at wire-speed performance
© 2016 Mellanox Technologies 18
NPS: Optimized for Data Plane
Parallel Performance
• Optimized for server architecture • Designed for highest single-thread performance • General purpose capabilities • Best for control plane and general purpose applications
• Optimized for networking systems • Designed for massive parallel processing • Large set of networking HW accelerations • Best for L2-L7 data plane processing
A single NPS-400 can deliver network processing capacity equal to 48 servers
(full rack) with power of less than one server Sing
le T
hrea
ded
Perf
orm
ance
© 2016 Mellanox Technologies 19
Data Center Network
Edge Router
TOR Switch
Rack of Servers
Appliances and Services Blades
An NFV network layer can be located in multiple network elements
© 2016 Mellanox Technologies 20
TOR Switch
Rack of Servers
Edge Router Appliances and Services Blades
Replace an ASIC-based TOR switch with a Smart NFV TOR: Network layer processing for VNFs Complete offload of vSwitch and network overlay Minimum server I/O overhead Significant reduction in cost and power
Smart NFV TOR
Smart NFV TOR Switch
© 2016 Mellanox Technologies 21
TOR Switch
Rack of Servers
Appliances and Services Blades
Edge Router
Smart edge router to facilitate an effective NFV architecture:
Deep packet classification for mapping traffic into flows and service chains
Dynamic load balancing among various distributed VNFs
Accommodate many subscribers and service chains
NFV Enabled Edge Router
NFV Enabled Edge Router
© 2016 Mellanox Technologies 22
TOR Switch
Rack of Servers
Blade or appliance with network processing resources:
Virtualized on-demand services per VNF
Servers and VNFs throughout the data center
Appliances and Services Blades
NFV Accelerated White Box/Appliances/Blades
Edge Router
NFV Service Blade/Appliance
© 2016 Mellanox Technologies 23
NPS in the Data Center
Smart NFV TOR
NFV Enabled Edge Router NFV Accelerated White Box/Appliances/Blades
Complete offload of vSwitch and network overlays
Network layer processing for VNFs
Mapping traffic into flows and service chains
Dynamic load balancing among distributed VMs & VNFs
© 2016 Mellanox Technologies 24
Summary
SDN & NFV virtualize the cloud network
Improve network agility, efficiency & cost
Network performance, scalability & management challenged
One NPS is equivalent to a rack of servers for network processing
NPS enables smart edge routers, TOR switches & appliances
For use by all network equipment and software vendors
Thank You