dynamic, latency-optimal vnf placement at the network edge

Dynamic, Latency-Optimal vNFPlacement at the Network Edge

Richard Cziva, Christos Anagnostopoulos, Dimitrios P Pezaros | University of Glasgow | [email protected] | INFOCOM’18 | Honolulu, Hawaii

Number of connected devices

Richard Cziva - Dynamic, Latency-Optimal vNF Placement at the Network Edge

Source: Ericsson IoT forecast https://www.ericsson.com/en/mobility-report/internet-of-things-forecast


Increased expectations


• Future networks are expected to support

� Context-aware

� Ultra-reliable

� User-specific network services

• Connected by

� High-bandwidth and

� Low-latency connections

Example services: video content caches, user-specific firewalls, DDoS mitigation modules, etc.


Opportunities with Edge NFVOne way to solve these challenges is to

bring Network Function Virtualization to the Network Edge


• Network Function Virtualization� Decoupling network services from

hardware and running them in software

� Used in data centers, in the core of the network

� Lacks latency-optimal service orchestration

• Multi-Access Edge Computing� Compute infrastructure at the edge

of the network� Also known as “fog computing”� Close proximity to the user => low

latency connectivity� Services at the edge save utilization

for the core


Edge NFVArchitecture



Latency-optimal vNF placement• We focus on placing vNFs to latency-optimal edge locations

� For each vNF association, we need to find a hosting device where a user-to-vNF end-to-end latency is minimal!

• Given: topology, hosting devices (with capabilities), latency on links, user’s locations

• Problem input: user to vNF assigment, vNF requirements (latency, compute)

• Output: vNF to edge mapping



NF (2CPU cores, max 30 ms from user)

Edge Node 1(2 CPU cores)

Router 5 ms

Edge Node 1(1 CPU core)5 ms2 ms

NF Edge Node 1Should be allocated to

Edge Node Edge Node

NFNF

Router

NFNF

Cloud DC

Router

InternetNF assignments

Edge properties:- CPU / Memory available

Link properties:- Bandwidth available (cost)- Delay

Goal is to have a placement, where:- All NFs are placed and traffic is

router through them- No overloading on links / edge

devices

While:Minimise latency of users

Delay: 40 ms

Delay: 40 ms

Delay: 5 ms Delay: 5 ms




Edge Node Edge Node

NF NF

Router

Cloud DC

Router

NF NF

Internet





devices


Delay: 40 ms

Delay: 40 ms





NFNF

NFNF

NF assignments

Edge vNF Placement ILP

Constraints

Objective function

Decision variable



Hardware limitations

Maximum latency

Allocate a vNF to 1 host

Bandwidth constraint

Valid path constraint

Are we done?• The ILP allocates vNFs to latency-optimal location. However:

� User’s move between edge devices� Latencies change on links frequently

� Other users impact traffic / congestion on the path

• These all impact the once optimal allocation!



We need dynamic re-allocation of edge vNFsto keep allocation latency-optimal!

Edge Node Edge Node

NF NF

Router

Cloud DC

Router

NF NF

Internet





devices


Delay: 40 ms

Delay: 40 ms





NFNF

NFNF

NF assignments

Edge Node Edge Node

NF NF

Router

Cloud DC

Router

NF NF

Internet





devices


Delay: 40 ms

Delay: 40 ms




Delay: 3 ms

User movement Optimal vNF allocation

Blue user moved between edge nodes-> allocation has to be re-evaluated

NFNF

NFNF

NF assignments

Edge Node Edge Node

NF NF

Router

Cloud DC

Router

NF NF

Internet

Delay: 40 ms

Delay: 40 ms


Edge NodeNo available capacity for

NFs

Router

Delay: 40 ms

Delay: 5 ms



What if the user moves further?

Delay: 3 ms

Edge Node Edge Node

NF NF

Router

Cloud DC

Router

NF NF

Internet

Delay: 40 ms

Delay: 40 ms



NFs

Router

Delay: 40 ms

Delay: 5 ms




User movement Sub-optimal vNF allocation

Edge Node Edge Node

Router

Cloud DC

Router

NF NF

Internet

Delay: 40 ms

Delay: 40 ms



NFs

Router

Delay: 40 ms

Delay: 5 ms




NF NF

User movement, vNF migrations Optimal vNF allocation

Latency violations• Assume that each vNF has a latency violation threshold that is a

maximum latency the vNF should get from the user. This is 𝜃I� For instance a cache vNF can have 20 ms for this value, while a control

plane vNF can have 150 ms

• Latency can not be guaranteed 100%, so the system will experience latency violations frequently

• Upcoming latency violations can be mitigated with a new latency-optimal vNF placement (but that costs migrations and placement calculation)



Goal: minimize latency violations, while keeping number of vNF migrations low

So, the new question is:How often (when) do we rearrange vNFs?



Every time we can� easy to implement,

always latency-optimal allocation

� way too many migrations

Periodically� easy to implement,

easy to predict migrations

� can results in too many latency violations, if the period is too long

Optimal time� low number of

latency violations and low number of migrations

How do we get this “optimal time”?• Counting latency violations experienced:

• The challenge is to find the (optimal stopping) time instance t* for deriving an optimal placement for the vNFs, such that Yt be as close to the system’s maximum tolerance Θ as possible



Migration cost between placements

Reward function Cumulative sum of all violations at time t

How do we get this “optimal time”?



Please find proof + solution fundamentals in the paper.

Note: we take only previous observations to make a decision.

Evaluation• We have divided the evaluation into two parts:

� Latency-optimal allocation� Placement scheduling (dynamic extension)

• Simulation environment:� Gurobi solver used for ILP (with Python binding)� Python implementation for the optimal stopping time triggering the solver

at the optimal stopping time



Edge vNF allocation



0

2

4

6

8

10

12

10 50 100 200 300 400 500 600 700 800 900 1000

Edge nodes reach resource capacity ->

Aver

age

late

ncy

from

use

rs to

thei

r vN

Fs (m

s)Number of total users (3 vNFs per user)

Allocating vNFs to edge devices and Cloud DCsAllocating vNFs to Cloud DCs only

Latency fluctuations



0

0.01

0.02

0.03

0.04

0.05

0.06

0.07

0.08

0 10 20 30 40 50 60 70 80 90 100

Prob

abilit

y

Latency (in ms)

k = 2.2, θ = 0.22

0

5

10

15

20

25

30

35

0 10 20 30 40 50 60 70 80 90 100

Late

ncy

(in m

s)

Time

Based on empirical data collected with Ruru.

Deviation from optimal



0

200

400

600

800

1000

1200

1400

1600

1800

0 10 20 30 40 50 60 70 80 90 100

Obj

ectiv

e fu

nctio

n (m

s)

Time instance

t=0 optimal allocationObjective function at t

0

1

2

3

4

5

6

0 10 20 30 40 50 60 70 80 90 100

Num

ber o

f vio

latio

ns

Time instance

Number of latency violations caused

Placement scheduling



Our solution does not reach the latency violation threshold, and gives low number of migrations.

0%

25%

50%

75%

100%

125%

0 100 200 300 400 500 600 700 800

Num

ber o

f cum

ulat

ive

viol

atio

ns to

war

ds th

e th

resh

old

Time

Migrating every time instanceOur scheduler based on optimal stopping time (our solution)

Our scheduler using P based on normal distributionMigrating every 100 time instance

Maximum number of violations tolerated Θ

0

10

100

1000

10000

0 100 200 300 400 500 600 700 800

Num

ber o

f cum

ulat

ive m

igra

tions

Time

Migrating every time instanceOur scheduler based at optimal stopping time (our solution)

Our scheduler using P based on normal distributionMigrating every 100 time instance

Summary• Edge vNFs can support low-latency – if allocated to the right devices

• Our work proposed a dynamic, latency-optimal vNF allocation algorithm� Optimal allocation used Integer Linear Programming� Dynamic extension was built on top of Optimal Stopping Theory

• Evaluation was conducted using real-world latency characteristics and a nation-wide network topology

• Our solution reduces the number of migrations by 94.8% and 76.9% compared to a scheduler that runs every time instance and one that would periodically trigger vNF migrations to a new optimal placement, respectively.



Thank you for your attention!Download this presentation from http://netlab.dcs.gla.ac.uk


Extra: learning phase



0

0.2

0.4

0.6

0.8

1

-1 0 1 2 3 4 5 6

PDF

Number of latency violations

Learned (actual) PP based on N(µ = 2.5, σ2 = 0.5)

0%

25%

50%

75%

100%

125%

0 100 200 300 400 500 600 700 800

Num

ber o

f cum

ulat

ive v

iola

tions

towa

rds

the

thre

shol

d

Time

Cum. number of violations using optimal stopping time (our solution)Maximum number of violations tolerated Θ

Glasgow Network Functions



dynamic, latency-optimal vnf placement at the network edge

Documents