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
Richard Cziva, Christos Anagnostopoulos, Dimitrios P Pezaros | University of Glasgow | [email protected] | INFOCOM’18 | Honolulu, Hawaii
Increased expectations
Richard Cziva - Dynamic, Latency-Optimal vNF Placement at the Network Edge
• 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.
Richard Cziva, Christos Anagnostopoulos, Dimitrios P Pezaros | University of Glasgow | [email protected] | INFOCOM’18 | Honolulu, Hawaii
Opportunities with Edge NFVOne way to solve these challenges is to
bring Network Function Virtualization to the Network Edge
Richard Cziva - Dynamic, Latency-Optimal vNF Placement at 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
Richard Cziva, Christos Anagnostopoulos, Dimitrios P Pezaros | University of Glasgow | [email protected] | INFOCOM’18 | Honolulu, Hawaii
Edge NFVArchitecture
Richard Cziva - Dynamic, Latency-Optimal vNF Placement at the Network Edge
Richard Cziva, Christos Anagnostopoulos, Dimitrios P Pezaros | University of Glasgow | [email protected] | INFOCOM’18 | Honolulu, Hawaii
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
Richard Cziva - Dynamic, Latency-Optimal vNF Placement at the Network Edge
Richard Cziva, Christos Anagnostopoulos, Dimitrios P Pezaros | University of Glasgow | [email protected] | INFOCOM’18 | Honolulu, Hawaii
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
Richard Cziva - Dynamic, Latency-Optimal vNF Placement at the Network Edge
Richard Cziva, Christos Anagnostopoulos, Dimitrios P Pezaros | University of Glasgow | [email protected] | INFOCOM’18 | Honolulu, Hawaii
Delay: 3 ms Delay: 3 ms
Edge Node Edge Node
NF NF
Router
Cloud DC
Router
NF NF
Internet
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
Richard Cziva - Dynamic, Latency-Optimal vNF Placement at the Network Edge
Richard Cziva, Christos Anagnostopoulos, Dimitrios P Pezaros | University of Glasgow | [email protected] | INFOCOM’18 | Honolulu, Hawaii
Delay: 3 ms Delay: 3 ms
NFNF
NFNF
NF assignments
Edge vNF Placement ILP
Constraints
Objective function
Decision variable
Richard Cziva - Dynamic, Latency-Optimal vNF Placement at the Network Edge
Richard Cziva, Christos Anagnostopoulos, Dimitrios P Pezaros | University of Glasgow | [email protected] | INFOCOM’18 | Honolulu, Hawaii
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!
Richard Cziva - Dynamic, Latency-Optimal vNF Placement at the Network Edge
Richard Cziva, Christos Anagnostopoulos, Dimitrios P Pezaros | University of Glasgow | [email protected] | INFOCOM’18 | Honolulu, Hawaii
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
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
Richard Cziva - Dynamic, Latency-Optimal vNF Placement at the Network Edge
Richard Cziva, Christos Anagnostopoulos, Dimitrios P Pezaros | University of Glasgow | [email protected] | INFOCOM’18 | Honolulu, Hawaii
Delay: 3 ms Delay: 3 ms
NFNF
NFNF
NF assignments
Edge Node Edge Node
NF NF
Router
Cloud DC
Router
NF NF
Internet
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
Richard Cziva - Dynamic, Latency-Optimal vNF Placement at the Network Edge
Richard Cziva, Christos Anagnostopoulos, Dimitrios P Pezaros | University of Glasgow | [email protected] | INFOCOM’18 | Honolulu, Hawaii
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
Delay: 5 ms Delay: 5 ms
Edge NodeNo available capacity for
NFs
Router
Delay: 40 ms
Delay: 5 ms
Richard Cziva - Dynamic, Latency-Optimal vNF Placement at the Network Edge
Richard Cziva, Christos Anagnostopoulos, Dimitrios P Pezaros | University of Glasgow | [email protected] | INFOCOM’18 | Honolulu, Hawaii
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
Delay: 5 ms Delay: 5 ms
Edge NodeNo available capacity for
NFs
Router
Delay: 40 ms
Delay: 5 ms
Richard Cziva - Dynamic, Latency-Optimal vNF Placement at the Network Edge
Richard Cziva, Christos Anagnostopoulos, Dimitrios P Pezaros | University of Glasgow | [email protected] | INFOCOM’18 | Honolulu, Hawaii
Delay: 3 ms Delay: 3 ms
User movement Sub-optimal vNF allocation
Edge Node Edge Node
Router
Cloud DC
Router
NF NF
Internet
Delay: 40 ms
Delay: 40 ms
Delay: 5 ms Delay: 5 ms
Edge NodeNo available capacity for
NFs
Router
Delay: 40 ms
Delay: 5 ms
Richard Cziva - Dynamic, Latency-Optimal vNF Placement at the Network Edge
Richard Cziva, Christos Anagnostopoulos, Dimitrios P Pezaros | University of Glasgow | [email protected] | INFOCOM’18 | Honolulu, Hawaii
Delay: 3 ms Delay: 3 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)
Richard Cziva - Dynamic, Latency-Optimal vNF Placement at the Network Edge
Richard Cziva, Christos Anagnostopoulos, Dimitrios P Pezaros | University of Glasgow | [email protected] | INFOCOM’18 | Honolulu, Hawaii
Goal: minimize latency violations, while keeping number of vNF migrations low
So, the new question is:How often (when) do we rearrange vNFs?
Richard Cziva - Dynamic, Latency-Optimal vNF Placement at the Network Edge
Richard Cziva, Christos Anagnostopoulos, Dimitrios P Pezaros | University of Glasgow | [email protected] | INFOCOM’18 | Honolulu, Hawaii
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
Richard Cziva - Dynamic, Latency-Optimal vNF Placement at the Network Edge
Richard Cziva, Christos Anagnostopoulos, Dimitrios P Pezaros | University of Glasgow | [email protected] | INFOCOM’18 | Honolulu, Hawaii
Migration cost between placements
Reward function Cumulative sum of all violations at time t
How do we get this “optimal time”?
Richard Cziva - Dynamic, Latency-Optimal vNF Placement at the Network Edge
Richard Cziva, Christos Anagnostopoulos, Dimitrios P Pezaros | University of Glasgow | [email protected] | INFOCOM’18 | Honolulu, Hawaii
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
Richard Cziva - Dynamic, Latency-Optimal vNF Placement at the Network Edge
Richard Cziva, Christos Anagnostopoulos, Dimitrios P Pezaros | University of Glasgow | [email protected] | INFOCOM’18 | Honolulu, Hawaii
Edge vNF allocation
Richard Cziva - Dynamic, Latency-Optimal vNF Placement at the Network Edge
Richard Cziva, Christos Anagnostopoulos, Dimitrios P Pezaros | University of Glasgow | [email protected] | INFOCOM’18 | Honolulu, Hawaii
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
Richard Cziva - Dynamic, Latency-Optimal vNF Placement at the Network Edge
Richard Cziva, Christos Anagnostopoulos, Dimitrios P Pezaros | University of Glasgow | [email protected] | INFOCOM’18 | Honolulu, Hawaii
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
Richard Cziva - Dynamic, Latency-Optimal vNF Placement at the Network Edge
Richard Cziva, Christos Anagnostopoulos, Dimitrios P Pezaros | University of Glasgow | [email protected] | INFOCOM’18 | Honolulu, Hawaii
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
Richard Cziva - Dynamic, Latency-Optimal vNF Placement at the Network Edge
Richard Cziva, Christos Anagnostopoulos, Dimitrios P Pezaros | University of Glasgow | [email protected] | INFOCOM’18 | Honolulu, Hawaii
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.
Richard Cziva - Dynamic, Latency-Optimal vNF Placement at the Network Edge
Richard Cziva, Christos Anagnostopoulos, Dimitrios P Pezaros | University of Glasgow | [email protected] | INFOCOM’18 | Honolulu, Hawaii
Thank you for your attention!Download this presentation from http://netlab.dcs.gla.ac.uk
Richard Cziva, Christos Anagnostopoulos, Dimitrios P Pezaros | University of Glasgow | [email protected] | INFOCOM’18 | Honolulu, Hawaii
Extra: learning phase
Richard Cziva - Dynamic, Latency-Optimal vNF Placement at the Network Edge
Richard Cziva, Christos Anagnostopoulos, Dimitrios P Pezaros | University of Glasgow | [email protected] | INFOCOM’18 | Honolulu, Hawaii
0
0.2
0.4
0.6
0.8
1
-1 0 1 2 3 4 5 6
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
Richard Cziva - Dynamic, Latency-Optimal vNF Placement at the Network Edge
Richard Cziva, Christos Anagnostopoulos, Dimitrios P Pezaros | University of Glasgow | [email protected] | INFOCOM’18 | Honolulu, Hawaii