Transcript Data Plane: Support Heterogeneity

Tesseract
A 4D Network Control Plane
Carnegie Mellon University
Microsoft Research
Rice University
Presented by: Alberto Gonzalez, Whitney Young
Current Designs


No direct control
Subtle dependencies

Example: load balance forwarding by tuning
OSPF link weights, but impacts inter-domain
routing
4D Architecture
Control plane:




Decision
Dissemination
Discovery
Data
Services:


Dissemination
Node configuration
Design
Design Goals







Timely reaction to network changes
Resilient to decision plane failure
Robust and secure control channels
Minimal switch configuration
Backward compatibility
Support diverse decision algorithms
Support multiple data planes
Implementation Overview

Switch


Implements data plane
Decision Element (DE)

Implements discovery, dissemination, and
decision planes
Decision Plane

Any network control algorithm can be easily
integrated





Incremental shortest path first
Spanning tree
Joint packet filtering/routing
Link cost-based traffic engineering
Resiliency to DE failure

Hot standbys receiving heartbeats
Dissemination Plane


Goal: communication between DEs and
switches
DEs handle most of dissemination plane, but
switches help out



Path to destination handled by DE
Switches have separate queue and dissemination
packets have higher priority
Security (protects switches, info passed through
dissemination plane, and compromised DEs)
Discovery Plane



Goal: minimize manual configuration
Switches send HELLO messages
DEs handle instructing the switches on what
to do once active


Initiate eBGP session with outside world
Backward compatibility (bootstrapping end
hosts)

Discovery plane as DHCP proxy
Data Plane

Configured by decision plane



WriteTable exposed with simple interface to
provide configuration service to decision plane
Allows easy implementation of different services
Decision/Dissemination Interface


Function independently of each other
Only 3 functions used to interface between them
(2 more simply to improve performance)
Performance Evaluation

Single Link Failures

Switch& Regional Failures
• Link Flapping


10-hop to 12-hop change
Tesseract can handle network changes
Performance Evaluation


1347 nodes & 6244 edges
DE Computation Time



Worst Case: 151ms
99th percentile: 40ms
Bandwidth overhead


Worst Case: 4.4MB
90% of switched updated with new state
Performance Evaluation

Failover times
Applications

In enterprise network:



Computers both new routes & packet filter
placements
Loads into routers with no forbidden traffic
leaked
No human involvement once security policy is
specified
Ethernet

Key features




Widely implemented frame format
Support for broadcasting frames
Transparent address learning model
Tesseract keeps these properties.
Ethernet


Through point comparisons
Control Plane for TCP flows



Started at 570Mbps
Leveled at 280Mbps after a failure
Conventional RSTP Control Plane



Starts at 280Mbps
Hit zero after failure
Recovered after 7-8 seconds at ~180Mbps
Summary

Tesseract

Robust


Secure


Convergence & Throughput
Scalable


Ethernet or IP
Good Performance


Enterprise Network
Resuable


Decission/Dissemination Planes
1,000+ Switches
Enables direct Control

Easier to Understand and Deploy