Virtual Middleboxes as

First-Class Entities in the Cloud

Aaron Gember, Anand Krishnamurthy,

Saul St. John, and Aditya Akella

1

Deploying Middleboxes Today

• Elastic compute and storage

• Limited add-on middlebox services

• Middlebox vendors are providing VM images

2

App VM

App VM

App VM

• Manually piece together tunnels and traffic splitters

– Brittle and inflexible

• Goal: low complexity for tenants

Challenge: Integration

3

• Scale based on CPU/memory

– Network significantly impacts performance

• Goal: high performance

Challenge: Performance

4

• Distributed placement

– Over-utilized network and under-utilized compute

• Goals: minimize costs; support more tenants

Challenge: Efficiency

5

& naïve flow distribution

Stratos (a.k.a. CloudMB)

6

Network-aware orchestration layer for middleboxes in clouds

VM Manager Placement

Flow Distribution Software Switches

100

250 470

360 680 730

Scaling

Stratos Scaling

• Based on application performance

• Scaling trials along a sequence

– If ↓ Latency: Keep and try another

– Else: Discard and move on

7

App

500 ms 400 ms 395 ms

Stratos Flow Distribution

• Minimizes inter-rack traffic using linear program

• Triggered following scaling, and periodically

8

Demonstration

9

Network-aware orchestration layer for middleboxes in clouds

• Deployable by cloud providers or third parties

• Minimizes network interactions

• Maximizes efficiency for tenants and providers

Stratos Summary

10

Visit us in the exhibit hall!

http://stratos.cs.wisc.edu