www.thalesgroup.com

MEC clusters great again! Geo-partitioning of MEC resources

Mathieu Bouet, Vania Conan Thales Communications & Security, France

2

This

do

cu

me

nt

ma

y n

ot

be

re

pro

du

ce

d, m

od

ifie

d,

ad

ap

ted

, p

ub

lish

ed

, tr

an

sla

ted

, in

an

y w

ay, in

wh

ole

or

in

pa

rt o

r d

isc

lose

d t

o a

th

ird

pa

rty w

ith

ou

t th

e p

rior

writt

en

co

nse

nt

of

Tha

les

- ©

Th

ale

s 2

01

5 A

ll rig

hts

re

serv

ed

.

ETSI MEC (Mobile Multi-access Edge Computing)

▌ Launched in Sept. 2014

▌ Key challenges:

Convergence between IT and Telecom (virtualization)

→ Elasticity and flexibility

Deploying various services and caching content at the mobile network edge

→ Reduced latency and core traffic

Allowing software applications to tap into local content and real-time information about local-access network conditions

→ Efficient resource management

Source: ETSI MEC

3

This

do

cu

me

nt

ma

y n

ot

be

re

pro

du

ce

d, m

od

ifie

d,

ad

ap

ted

, p

ub

lish

ed

, tr

an

sla

ted

, in

an

y w

ay, in

wh

ole

or

in

pa

rt o

r d

isc

lose

d t

o a

th

ird

pa

rty w

ith

ou

t th

e p

rior

writt

en

co

nse

nt

of

Tha

les

- ©

Th

ale

s 2

01

5 A

ll rig

hts

re

serv

ed

.

High level view of a MEC deployment

▌Applications:

Live video analysis

Privacy filter

Personal assistant (productivity, sport…)

Remote medicine

…

▌ Leverage virtualization at the edge

▌N-level hierarchy

4

This

do

cu

me

nt

ma

y n

ot

be

re

pro

du

ce

d, m

od

ifie

d,

ad

ap

ted

, p

ub

lish

ed

, tr

an

sla

ted

, in

an

y w

ay, in

wh

ole

or

in

pa

rt o

r d

isc

lose

d t

o a

th

ird

pa

rty w

ith

ou

t th

e p

rior

writt

en

co

nse

nt

of

Tha

les

- ©

Th

ale

s 2

01

5 A

ll rig

hts

re

serv

ed

.

Context and motivation

▌ Spatiotemporal variations

Mobile communications are generally spatially distributed according to the population density and activity, which vary in time

▌Activity patterns

The mobile traffic in the business areas differ from the mobile traffic in the transport, residential and entertainment areas [10, 14, 17]

▌ Such properties will be amplified with the realization of the IoT and 5G visions [7]

5

This

do

cu

me

nt

ma

y n

ot

be

re

pro

du

ce

d, m

od

ifie

d,

ad

ap

ted

, p

ub

lish

ed

, tr

an

sla

ted

, in

an

y w

ay, in

wh

ole

or

in

pa

rt o

r d

isc

lose

d t

o a

th

ird

pa

rty w

ith

ou

t th

e p

rior

writt

en

co

nse

nt

of

Tha

les

- ©

Th

ale

s 2

01

5 A

ll rig

hts

re

serv

ed

.

Objective: Dimension MEC areas (or clusters)

▌ A MEC partitioning must have the following properties

MEC servers, as any compute, storage and network node, have a maximum capacity (e.g. in terms of CPU, storage resources, application hosting capabilities etc.)

MEC server loads should be balanced both spatially and temporally to improve user experience

The traffic between the MEC servers and the core should be minimized, in particular by consolidating applications at the MEC server level, such that the global latency is reduced

A MEC cluster should be geo-consistent (connected) to rationalize the deployment

6

This

do

cu

me

nt

ma

y n

ot

be

re

pro

du

ce

d, m

od

ifie

d,

ad

ap

ted

, p

ub

lish

ed

, tr

an

sla

ted

, in

an

y w

ay, in

wh

ole

or

in

pa

rt o

r d

isc

lose

d t

o a

th

ird

pa

rty w

ith

ou

t th

e p

rior

writt

en

co

nse

nt

of

Tha

les

- ©

Th

ale

s 2

01

5 A

ll rig

hts

re

serv

ed

.

Related work

▌Partitioning and MEC server placement

Qazi [11] showed that the number and the locations of MEC servers have a direct

impact on the QoE (imbalance loads and high latencies) and on the operational

cost

Ceselli et al. [19] have proposed an ILP for the joint problem of base stations

allocation to MEC servers and routing to reduce infrastructure cost. The clusters

are not geo-consistent, meaning that the base stations associated to a MEC

server can be completely scattered in space. The computation does not scale.

And us!

[11] Z. Ayyub Qazi, P. Krishna, V. Sekar, V. Gopalakrishnan, K. Joshi, and S. Das. 2016. KLEIN: A Minimally Disruptive Design for an Elastic Cellular Core. In Proceedings of ACM Symposium on SDN Research (SOSR). [19] A. Ceselli, M. Premoli, and S. Secci. 2017. Mobile Edge Cloud Network Design Optimization. IEEE/ACM Transactions on Networking 99 (2017), 1–14.

7

This

do

cu

me

nt

ma

y n

ot

be

re

pro

du

ce

d, m

od

ifie

d,

ad

ap

ted

, p

ub

lish

ed

, tr

an

sla

ted

, in

an

y w

ay, in

wh

ole

or

in

pa

rt o

r d

isc

lose

d t

o a

th

ird

pa

rty w

ith

ou

t th

e p

rior

writt

en

co

nse

nt

of

Tha

les

- ©

Th

ale

s 2

01

5 A

ll rig

hts

re

serv

ed

.

Our geo-partitioning algorithm for MEC resources

▌MEC clustering algorithm inspired from the Louvain method (detection of

communities in graphs)

▌Aggregates local interactions (communications) up to a max. cluster load

communication MEC cluster

?

8

This

do

cu

me

nt

ma

y n

ot

be

re

pro

du

ce

d, m

od

ifie

d,

ad

ap

ted

, p

ub

lish

ed

, tr

an

sla

ted

, in

an

y w

ay, in

wh

ole

or

in

pa

rt o

r d

isc

lose

d t

o a

th

ird

pa

rty w

ith

ou

t th

e p

rior

writt

en

co

nse

nt

of

Tha

les

- ©

Th

ale

s 2

01

5 A

ll rig

hts

re

serv

ed

.

Algorithm: Initialization

comm.

9

This

do

cu

me

nt

ma

y n

ot

be

re

pro

du

ce

d, m

od

ifie

d,

ad

ap

ted

, p

ub

lish

ed

, tr

an

sla

ted

, in

an

y w

ay, in

wh

ole

or

in

pa

rt o

r d

isc

lose

d t

o a

th

ird

pa

rty w

ith

ou

t th

e p

rior

writt

en

co

nse

nt

of

Tha

les

- ©

Th

ale

s 2

01

5 A

ll rig

hts

re

serv

ed

.

Algorithm: 1st pass

The two grid cells that have the highest interaction are merged

if it respects the max cluster capacity

10

This

do

cu

me

nt

ma

y n

ot

be

re

pro

du

ce

d, m

od

ifie

d,

ad

ap

ted

, p

ub

lish

ed

, tr

an

sla

ted

, in

an

y w

ay, in

wh

ole

or

in

pa

rt o

r d

isc

lose

d t

o a

th

ird

pa

rty w

ith

ou

t th

e p

rior

writt

en

co

nse

nt

of

Tha

les

- ©

Th

ale

s 2

01

5 A

ll rig

hts

re

serv

ed

.

Algorithm: 2nd pass

The two grid cells that have the highest interaction are merged

if it respects the max cluster capacity

Until no pair of

nodes can be

merged

because of

max. server

capacity

11

This

do

cu

me

nt

ma

y n

ot

be

re

pro

du

ce

d, m

od

ifie

d,

ad

ap

ted

, p

ub

lish

ed

, tr

an

sla

ted

, in

an

y w

ay, in

wh

ole

or

in

pa

rt o

r d

isc

lose

d t

o a

th

ird

pa

rty w

ith

ou

t th

e p

rior

writt

en

co

nse

nt

of

Tha

les

- ©

Th

ale

s 2

01

5 A

ll rig

hts

re

serv

ed

.

Dataset (1/2)

▌ In 2014, Telecom Italia organized the ‘Telecom Italia Big Data Challenge’ (http://www.telecomitalia.com/tit/en/bigdatachallenge/contest.html)

▌ Several types of Call Details Record (CDR) datasets were produced to measure the interaction intensity between different locations

▌ The dataset we used in this study:

quantify the interactions within Milan (i.e., Milan to Milan) over November 2013

temporally aggregated every 10 min and spatially aggregated in a grid (next slide)

▌ (at most) 34% of population's data is collected, due to Telecom Italia's market share. Moreover there is no information about missed calls.

Gianni Barlacchi et al., «A multi-source dataset of urban life in the city of Milan and the Province of Trentino”, in Science Data 2015

12

This

do

cu

me

nt

ma

y n

ot

be

re

pro

du

ce

d, m

od

ifie

d,

ad

ap

ted

, p

ub

lish

ed

, tr

an

sla

ted

, in

an

y w

ay, in

wh

ole

or

in

pa

rt o

r d

isc

lose

d t

o a

th

ird

pa

rty w

ith

ou

t th

e p

rior

writt

en

co

nse

nt

of

Tha

les

- ©

Th

ale

s 2

01

5 A

ll rig

hts

re

serv

ed

.

Calls intensity

City map of Milan Spatial discretization of Milan

area (d=235m)

Dataset (2/2)

Call intensity: number proportional to the number of calls generated from one grid cell to one other grid cell

Normalized mobile

communication intensity

(5pm-6pm, 11/04/2013)

13

This

do

cu

me

nt

ma

y n

ot

be

re

pro

du

ce

d, m

od

ifie

d,

ad

ap

ted

, p

ub

lish

ed

, tr

an

sla

ted

, in

an

y w

ay, in

wh

ole

or

in

pa

rt o

r d

isc

lose

d t

o a

th

ird

pa

rty w

ith

ou

t th

e p

rior

writt

en

co

nse

nt

of

Tha

les

- ©

Th

ale

s 2

01

5 A

ll rig

hts

re

serv

ed

.

Evaluation - Varying day and day time (1/2)

Number of MEC clusters

▌ Logically, as the maximum cluster capacity

diminishes the number of clusters increases to serve traffic at the edge

14

This

do

cu

me

nt

ma

y n

ot

be

re

pro

du

ce

d, m

od

ifie

d,

ad

ap

ted

, p

ub

lish

ed

, tr

an

sla

ted

, in

an

y w

ay, in

wh

ole

or

in

pa

rt o

r d

isc

lose

d t

o a

th

ird

pa

rty w

ith

ou

t th

e p

rior

writt

en

co

nse

nt

of

Tha

les

- ©

Th

ale

s 2

01

5 A

ll rig

hts

re

serv

ed

.

Evaluation - Varying day and day time (1/2)

Number of MEC clusters

▌ Logically, as the maximum cluster capacity

diminishes the number of clusters increases to serve traffic at the edge

Core offloaded

Intra MEC cluster vs. total traffic (%)

▌ Traffic more localized on week-end

15

This

do

cu

me

nt

ma

y n

ot

be

re

pro

du

ce

d, m

od

ifie

d,

ad

ap

ted

, p

ub

lish

ed

, tr

an

sla

ted

, in

an

y w

ay, in

wh

ole

or

in

pa

rt o

r d

isc

lose

d t

o a

th

ird

pa

rty w

ith

ou

t th

e p

rior

writt

en

co

nse

nt

of

Tha

les

- ©

Th

ale

s 2

01

5 A

ll rig

hts

re

serv

ed

.

Evaluation - Varying day and day time (2/2)

Clustering result

For a maximum cluster capacity of 5% of the total communications, i.e. 8,500 communications (5pm-6pm, 11/04/2013). Th‡e numbers in the clusters correspond to

their load.

Well balanced server load

MEC cluster loads

▌ The median is close to the maximum cluster capacity

16

This

do

cu

me

nt

ma

y n

ot

be

re

pro

du

ce

d, m

od

ifie

d,

ad

ap

ted

, p

ub

lish

ed

, tr

an

sla

ted

, in

an

y w

ay, in

wh

ole

or

in

pa

rt o

r d

isc

lose

d t

o a

th

ird

pa

rty w

ith

ou

t th

e p

rior

writt

en

co

nse

nt

of

Tha

les

- ©

Th

ale

s 2

01

5 A

ll rig

hts

re

serv

ed

.

Evaluation – Through time (1/2)

▌ Traffic offloaded to the core (i.e. cluster saturation) < 3%

▌ Intra-core traffic ratio around 53%

Static clusters + dynamic demand => almost no server saturation!

17

This

do

cu

me

nt

ma

y n

ot

be

re

pro

du

ce

d, m

od

ifie

d,

ad

ap

ted

, p

ub

lish

ed

, tr

an

sla

ted

, in

an

y w

ay, in

wh

ole

or

in

pa

rt o

r d

isc

lose

d t

o a

th

ird

pa

rty w

ith

ou

t th

e p

rior

writt

en

co

nse

nt

of

Tha

les

- ©

Th

ale

s 2

01

5 A

ll rig

hts

re

serv

ed

.

Evaluation – Through time (2/2)

Well balanced server load through time

▌ Normalized MEC cluster loads over a day

(11/04/2015) with a partition done at 5pm and a

maximum cluster capacity of 5% of the total

communications

18

This

do

cu

me

nt

ma

y n

ot

be

re

pro

du

ce

d, m

od

ifie

d,

ad

ap

ted

, p

ub

lish

ed

, tr

an

sla

ted

, in

an

y w

ay, in

wh

ole

or

in

pa

rt o

r d

isc

lose

d t

o a

th

ird

pa

rty w

ith

ou

t th

e p

rior

writt

en

co

nse

nt

of

Tha

les

- ©

Th

ale

s 2

01

5 A

ll rig

hts

re

serv

ed

.

Conclusion

▌We proposed a graph-based geo-partitioning algorithm for MEC resources

▌ The data-driven evaluation shows

Core offloading (i.e. consolidation of the traffic at the edge)

Well balanced server loads (even through time)

▌ Future work

Mathematical optimization model

Group communications

Combination with online application offloading and migration

Experiments with SDN and NFV

19

This

do

cu

me

nt

ma

y n

ot

be

re

pro

du

ce

d, m

od

ifie

d,

ad

ap

ted

, p

ub

lish

ed

, tr

an

sla

ted

, in

an

y w

ay, in

wh

ole

or

in

pa

rt o

r d

isc

lose

d t

o a

th

ird

pa

rty w

ith

ou

t th

e p

rior

writt

en

co

nse

nt

of

Tha

les

- ©

Th

ale

s 2

01

5 A

ll rig

hts

re

serv

ed

.

References

▌ [1] 2014. Open Big Data. httŠps://dandelion.eu/datamine/open-big-data/. (2014). Accessed: 2017-03-23.

▌ [2] G. Barlacchi, M De Nadai, R Larcher, and others. 2015. A multi-source dataset of urban life in the city of Milan and the Province of Trentino. Scienti€c Data 2, 150055 (2015).

▌ [3] V.D. Blondel, J.L. Guillaume, R. LambioŠe, and E.L.J.S. Mech. 2008. Fast unfolding of communities in large networks. J. Stat. Mech (2008).

▌ [4] C.-Y. Chang, K. Alexandris, N. Nikaein, K. Katsalis, and T. Spyropoulos. 2016. MEC architectural implications for LTE/LTE-A networks. In Proceedings of the Workshop on Mobility in the Evolving Internet Architecture (MobiArch). 13–18.

▌ [5] European Telecommunications Standards Institute (ETSI). 2015. Mobile-Edge Computing (MEC); Service Scenarios (GS MEC-IEG 004).

▌ [6] OpenFog Consortium Architecture Working Group. 2016. OpenFog Architecture Overview (OPFWP001.0216 ).

▌ [7] Y. C. Hu, M. Patel, D. Sabella, N. Sprecher, and V. Young. 2015. Mobile Edge Computing A key technology towards 5G (ETSI White Paper No. 11). European

Telecommunications Standards Institute (ETSI).

▌ [8] F. Manco, J. Martins, K. Yasukata, J. Mendes, S. Kuenzer, and F. Huici. 2015. ŒThe Case for the Superƒuid Cloud. In In Proceedings of USENIX Workshop on Hot Topics in Cloud Computing (HotCloud).

▌ [9] Y. Mao, C. You, J. Zhang, K. Huang, and K. Ben Letaief. 2017. Mobile Edge Computing: Survey and Research Outlook. CoRR abs/1701.01090 (2017). hŠttp//arxiv.org/abs/1701.01090

▌ [10] D. Naboulsi, M. Fiore, S. Ribot, and R. Stanica. 2015. Large-scale Mobile Traffic Analysis: a Survey. IEEE Communications Surveys and Tutorials (2015).

▌ [11] Z. Ayyub Qazi, P. Krishna, V. Sekar, V. Gopalakrishnan, K. Joshi, and S. Das. 2016. KLEIN: A Minimally Disruptive Design for an Elastic Cellular Core. In

Proceedings of ACM Symposium on SDN Research (SOSR).

▌ [12] R. Saunders, J. Cho, A. Banerjee, F. Rocha, and J. Van der Merwe. 2016. P2P Offloading in Mobile Networks using SDN. In Proceedings of ACM Symposium on SDN Research (SOSR).

▌ [13] H. Tan, Z. Han, X.Y. Li, and F.C.M. Lau. 2017. Online Job Dispatching and Scheduling in Edge-Clouds. In In Proceedings of IEEE International Conference on Computer Communications (INFOCOM). 1–9.

▌ [14] N. Tastevin and M. Bouet. 2016. Characterizing and modeling the distance of mobile calls: A metropolitan case study. In In Proceedings of IEEE 27th Annual International Symposium on Personal, Indoor, and Mobile Radio Communications (PIMRC).

▌ [15] L. Tong, Y. Li, and W. Gao. 2016. A hierarchical edge cloud architecture for mobile computing. In In Proceedings of IEEE International Conference on Computer Communications (INFOCOM). 1–9.

▌ [16] P.L. Ventre, C. Pisa, S. Salsano, G. Siracusano, F. Schmidt, P. Lungaroni, and N. Blefari-Melazzi. 2016. Performance Evaluation and Tuning of Virtual Infrastructure Managers for (Micro) Virtual Network Functions. In In Proceedings of IEEE

Conference on Network unction Virtualization and Software Defined Networks (NFV-SDN).

▌ [17] H. Wang, F. Xu, Y. Li, P. Zhang, and D. Jin. 2015. Understanding Mobile Traffic Patterns of Large Scale Cellular Towers in Urban Environment. In Proceedings of ACM Internet Measurement Conference (IMC).

▌ [18] Q. Xu, F. Qian, J. Huang, A. Gerber, Z. Wang, and Z. M. Mao. 2011. Cellular data network infrastructure characterization and implication on mobile content placement. In Proceedings of ACM SIGMETRICS.

▌ [19] A. Ceselli, M. Premoli, and S. Secci. 2017. Mobile Edge Cloud Network Design Optimization. IEEE/ACM Transactions on Networking 99 (2017), 1–14.