PdP: Parallelizing Data Plane in Virtual Network Substrate Yong - - PowerPoint PPT Presentation

PdP: Parallelizing Data Plane in Virtual Network Substrate

Yong Liao, Dong Yin, Lixin Gao University of Massachusetts at Amherst

Network Virtualization Platform

Multiple heterogeneous concurrent virtual networks

Flexibility

Customizable virtual networks

High-performance

Good forwarding speed

Isolation

Minimal interference

Low-cost

Facilitate wide-area deployment

` ` ` `

Existing Network Virtualization Platforms

VINI

User mode forwarding, slow, highly customizable

Trellis

Kernel mode forwarding, faster, less customizable

VRouter (Xen)

Close to native speed Needs hardware support to scale

Supercharging Planetlab

Special-purpose hardware, superior speed Harder to program

Existing Network Virtualization Platforms

Flexibility Performance Isolation Cost

VINI Good Slow forwarding Good Low cost Trellis Moderate Close to native speed Moderate Low cost VRouter Good Close to native speed Good High SPP Moderate Superior speed Moderate High

Main Ideas of PdP

Accelerate data forwarding with multiple

forwarding engines

Faster aggregate forwarding speed Commodity hardware is inexpensive

Run virtual network data plane and control

plane in VMs

Isolation among virtual networks Better flexibility to customization

Architecture of PdP

multiplexer & demultiplexer

incoming, unprocessed packets

• utgoing, processed

packets

management host (vnet control plane) forwarding engines (vnet data plane)

Multiple forwarding engines (FEs)

Sliced into virtual nodes Isolation

Multiplexer & demultiplexer

Classify packets to data plane

VMs

Send packets out to physical

NICs

High-speed

Control plane and data plane

running in VMs

Customizable Isolation and management

VNet Data Plane

Mapping between VNet and Forwarding Engines

Multiple FEs for one VNet How to allocate FEs to VNets

Each virtual node performs (in user mode)

Lookup, encapsulation for virtual links

Multiplexer & Demultiplexer

Packet classifier

Different ports for different VMs

Packet dispatcher sends packets out

FE already marked the outgoing NIC

Can potentially be bottleneck

Prototype Implementation

Commodity PCs

P4 2.6GHz CPU, 1G mem, Gbit

NIC

Multiplexer & demultiplexer

Kernel mode Click

VNet Data plane

User mode Click in VM

VNet Control plane

XORP in VM

Interaction between VNet

control plane and data plane

Updating forwarding table by

multicast

Gbit Ethernet Switch forwarding engines control plane host and multiplexer&demultiplexer

Raw UDP packet forwarding Speed

UDP packet forwarding speed 50 100 150 200 250 300 350 200 400 600 800 1000 1200

input speed (Kpps)

forwarding speed (Kpps)

user Click kernel Click

• ne FE

two FEs three FEs

Small table: two entries, Similar for large table

Raw UDP Packet Loss Rate and RTT

Loss rate in UDP packet forwarding

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 200 400 600 800 1000 1200 input speed (Kpps) loss rate

user Click kernel Click

• ne FE

two FEs three FEs

User Click PdP Kernel Click Two-hop RTT(ms) 0.208 0.296 0.132

TCP Performance

Aggregate throughput is close to kernel Click Out-of-order packets

TCP Throughput Experiment Results

360 369 565 763 860

100 200 300 400 500 600 700 800 900 1000 User Click One FE Two FEs Three Fes Kernel Click Throughput (Mbps)

round-robin proportional % of out-of-

• rder pkts

12.27% 10.02%

• ne FE

two FEs three FEs % of out-of-

• rder pkts

0.31% 10.19% 13.02%

Conclusion and Future Work

PdP provides the maximal flexibility to customize

VNets

Forwarding speed of PdP scales with the

number of FEs

Hardware multiplexer/demultiplexer Flow based classification (out-of-order packet

problem)