MON: MISSION-OPTIMIZED OVERLAY NETWORKS Bruce Spang , Anirudh - - PowerPoint PPT Presentation

▶

Mar 17, 2023 435 likes •683 views

MON: MISSION-OPTIMIZED OVERLAY NETWORKS Bruce Spang , Anirudh Sabnis, Ramesh Sitaraman, Don Towsley, Brian Decleene SETTING Managed by a single organization Good model of underlying network Di ff erent types of tra ffi c use this

SLIDE 1

MON: MISSION-OPTIMIZED OVERLAY NETWORKS

Bruce Spang, Anirudh Sabnis, Ramesh Sitaraman, Don Towsley, Brian Decleene

SLIDE 2

SLIDE 3

SETTING

➤ Managed by a single organization ➤ Good model of underlying network ➤ Different types of traffic use this network ➤ The organization cares more about certain types of traffic

SLIDE 4

MON

SLIDE 5

FLOWS

➤ A flow corresponds to traffic being routed between unique

pair of source-destination, traffic type (ex. VOIP , Video).

➤ A flow could use multiple overlay paths. ➤ A flow could have multiple connections. ➤ Example. 10 VOIP connections from BAE to UMass

SLIDE 6

UTILITY FUNCTIONS

➤ Utility Function is a measure of “usefulness” for a certain

rate (x Mbps) allocated for a single connection of the flow.

➤ Increasing function, possibly non-convex

SLIDE 7

MON OPTIMIZATION PROBLEM

➤ k - a flow ➤ nk - number of connections

➤ xk,r - rate of flow k on path r ➤ Uk(x) - Utility of flow k ➤ Cl - estimated capacity ➤ Nk - estimated demand

SLIDE 8

PFO

Solve the problem offline using Global Optimization Techniques

SLIDE 9

MOTIVATION FOR CONGESTION CONTROL

➤ Solving optimization problem may take time ➤ What if network changes? Or demand changes? ➤ Quickly react to packet loss and increased bandwidth ➤ We use TCP for this

SLIDE 10

MONTRA OPTIMIZATION MODEL

SLIDE 11

TCP - INCREASE/DECREASE RULES

➤ xf - sending rate of flow f ➤ γ - stability constant ➤ wf - weight

SLIDE 12

HOW TO PICK WF?

➤ Controllers should match PFO’s target rates ➤ Controllers should make good decisions if network changes

SLIDE 13

MAPPING PFO TO MONTRA

Theorem Idea: Assume PFO has an accurate model of the

network. Suppose PFO picks a target rate A. If the gradient of

MONtra matches the gradient of PFO at A, then MONtra will converge to A Proof: See the paper

SLIDE 14

MON

SLIDE 15

UTILITY FUNCTIONS AND FLOWS

➤ Used following utility functions: ➤ Generated random flows

Flow A Flow B

SLIDE 16

DOES MONTRA SEND AT THE CORRECT RATE?

MONtra converges to PFO’s rates on a complex network

r15 r16 r17 r21 r2 r13 s30 r9 r18 r19 r20 r22 s31 s32 s33 s34 r23 r24 s35 r12 s25 r6 s26 r7 s27 s28 s29 r14 r0 r1 r3 r8 r4 r5 r10 r11 Site Router

SLIDE 17

BENEFITS OF OVERLAY ROUTING

Number of hops: Mission utility increases with number

f hops

Number of paths: Allowing more paths increases utility

SLIDE 18

HOW DOES MONTRA REACT TO SLIGHT CHANGES IN NETWORK?

Ran PFO with x=3 Mbps
Adjusted capacity from 1 to 9 Mbps.
MONtra does better than baselines

x Mbps 10 Mbps

5 Mbps

SLIDE 19

HOW DOES MONTRA REACT TO NETWORK FAILURES?

Triangle topology:

Link fails PFO recovers

SLIDE 20

HOW DOES MONTRA REACT TO NETWORK FAILURES?

Link failure decreases utility
PFO recovers utility

AT&T topology:

SLIDE 21

SUMMARY

➤ Overlay network architecture ➤ Optimally route flows ➤ Resilient to changes to inputs

SLIDE 22

THANKS!

SLIDE 23

MON: MISSION-OPTIMIZED OVERLAY NETWORKS

THANKS!

QUESTIONS?