Virtual Network Embedding with Collocation Benefits and Limitations - - PowerPoint PPT Presentation

▶
virtual network embedding with collocation
SMART_READER_LITE
LIVE PREVIEW

Virtual Network Embedding with Collocation Benefits and Limitations - - PowerPoint PPT Presentation

Jan 04, 2024 238 likes •884 views

Virtual Network Embedding with Collocation Benefits and Limitations of Pre-Clustering urst 1 , Stefan Schmid 2 , Anja Feldmann 1 Carlo F 1: TU Berlin 2: TU Berlin & Telekom Innovation Laboratories November 12, 2013 Carlo F urst (TU

slide-1
SLIDE 1

Virtual Network Embedding with Collocation

Benefits and Limitations of Pre-Clustering Carlo F¨ urst1 , Stefan Schmid2, Anja Feldmann1

1: TU Berlin 2: TU Berlin & Telekom Innovation Laboratories

November 12, 2013

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 1 / 21

slide-2
SLIDE 2

Today’s Datacenters...

Multi-tenant virtualized Tenants typically pay for host resources Connectivity is guaranteed

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 2 / 21

slide-3
SLIDE 3

Today’s Datacenters...

Multi-tenant virtualized Tenants typically pay for host resources Connectivity is guaranteed

Problem [Ballani’11]:

Studies have shown that the intra-cloud bandwidth can vary by an order of magnitude. ⇒ Unpredictable application performance

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 2 / 21

slide-4
SLIDE 4

Remove the uncertainty ? ?

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 3 / 21

slide-5
SLIDE 5

Remove the uncertainty

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 3 / 21

slide-6
SLIDE 6

Outline

Explain model and problem Identify the impact of the collocation option on embedding algorithms Introduce Pre-Clustering - a technique to enable any existing algorithm to generate collocated embeddings

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 4 / 21

slide-7
SLIDE 7

Virtual Network Embedding Problem Physical Machine Physical Link

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 5 / 21

slide-8
SLIDE 8

Virtual Network Embedding Problem

  • Abstract aggregated

"Compute Resource"

  • Bandwidth

Physical Machine Physical Link

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 5 / 21

slide-9
SLIDE 9

Virtual Network Embedding Problem

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 5 / 21

slide-10
SLIDE 10

Virtual Network Embedding Problem Virtual Node Virtual Link

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 5 / 21

slide-11
SLIDE 11

Virtual Network Embedding Problem Virtual Node Virtual Link

  • Requested Compute Units
  • Requested Bandwidth

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 5 / 21

slide-12
SLIDE 12

Virtual Network Embedding Problem

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 5 / 21

slide-13
SLIDE 13

What is a ‘good’ mapping?

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 6 / 21

slide-14
SLIDE 14

What is a ‘good’ mapping?

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 6 / 21

slide-15
SLIDE 15

Existing Solutions

Many existing mapping algorithms ViNE [CHOWDHURY, Infocom 2009] SecondNet [GUO, Co-NEXT 2010] Oktopus [BALLANI, Sigcomm 2011] Isomorphism Detection [LISCHKA, Sigcomm 2009] Various Mixed-Integer-Programs . . .

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 7 / 21

slide-16
SLIDE 16

Existing Solutions

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 7 / 21

slide-17
SLIDE 17

Existing Solutions

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 7 / 21

slide-18
SLIDE 18

Collocated Mappings Physical Machine with capacity 2

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 8 / 21

slide-19
SLIDE 19

Collocated Mappings

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 8 / 21

slide-20
SLIDE 20

Benchmarking Algorithm: LoCo

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 9 / 21

slide-21
SLIDE 21

Benchmarking Algorithm: LoCo

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 9 / 21

slide-22
SLIDE 22

Benchmarking Algorithm: LoCo

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 9 / 21

slide-23
SLIDE 23

Benchmarking Algorithm: LoCo

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 9 / 21

slide-24
SLIDE 24

Benchmarking Algorithm: LoCo

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 9 / 21

slide-25
SLIDE 25

Benchmarking Algorithm: LoCo

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 9 / 21

slide-26
SLIDE 26

Benchmarking Algorithm: LoCo

Backtrack on failure Backtrack only over possible start nodes Graph exploration is directed by node / link resource requests Avoid Backtracking by forward checking

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 9 / 21

slide-27
SLIDE 27

Evaluation Setup

Request sequence ADD REQ1 ADD REQ2 ADD REQ3 REM REQ1 ADD REQ4 STATE ...

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 10 / 21

slide-28
SLIDE 28

Evaluation Setup

Request sequence ADD REQ1 ADD REQ2 ADD REQ3 REM REQ1 ADD REQ4 STATE ... Add Requests Until: Sum of requested node resources = Sum of substrate node resources

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 10 / 21

slide-29
SLIDE 29

Evaluation Setup

Request sequence ADD REQ1 ADD REQ2 ADD REQ3 REM REQ1 ADD REQ4 STATE ... Measure node utilization

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 10 / 21

slide-30
SLIDE 30

Evaluation Setup

Request sequence ADD REQ1 ADD REQ2 ADD REQ3 REM REQ1 ADD REQ4 STATE ... Increase time until a Request expires

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 10 / 21

slide-31
SLIDE 31

Evaluation Setup

Request sequence ADD REQ1 ADD REQ2 ADD REQ3 REM REQ1 ADD REQ4 STATE ... Add Requests Until: ...

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 10 / 21

slide-32
SLIDE 32

Evaluation Setup

Request sequence ADD REQ1 ADD REQ2 ADD REQ3 REM REQ1 ADD REQ4 STATE ... Substrate Topologies FatTree

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 10 / 21

slide-33
SLIDE 33

Evaluation Setup

Embed. Algorithm Request sequence ADD REQ1 ADD REQ2 ADD REQ3 REM REQ1 ADD REQ4 STATE ... Substrate Topologies FatTree Unmodified Requests LoCo SNet

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 10 / 21

slide-34
SLIDE 34

Impact of the collocation option

  • Loco

SecondNet 0.0 0.2 0.4 0.6 0.8 1.0

Slight Impact

Node Utilization

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 11 / 21

slide-35
SLIDE 35

Impact of the collocation option

  • Loco

SecondNet 0.0 0.2 0.4 0.6 0.8 1.0

Slight Impact

Node Utilization

  • Loco

SecondNet 0.0 0.2 0.4 0.6 0.8 1.0

Strong Impact

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 11 / 21

slide-36
SLIDE 36

Impact of the collocation option

  • Loco

SecondNet 0.0 0.2 0.4 0.6 0.8 1.0

Slight Impact

Node Utilization

  • Loco

SecondNet 0.0 0.2 0.4 0.6 0.8 1.0

Strong Impact

  • Loco

SecondNet 0.0 0.2 0.4 0.6 0.8 1.0

Average Impact

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 11 / 21

slide-37
SLIDE 37

Can we leverage the benefits of collocation with the existing algorithms?

slide-38
SLIDE 38

Pre-Clustering

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 13 / 21

slide-39
SLIDE 39

Pre-Clustering

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 13 / 21

slide-40
SLIDE 40

Pre-Clustering 4 1 3 4

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 13 / 21

slide-41
SLIDE 41

Pre-Clustering

We use: Farhat LoCo OptCut (runtime optimized MIP)

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 13 / 21

slide-42
SLIDE 42

LoCo Preclustering

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 14 / 21

slide-43
SLIDE 43

LoCo Preclustering

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 14 / 21

slide-44
SLIDE 44

LoCo Preclustering 2

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 14 / 21

slide-45
SLIDE 45

OptCut

Generates an optimal (w.r.t. the amount of link resources between the merged nodes) Pre-Clustering Substrate is represented by two numbers:

◮ MAXV : The estimated host resources of a node ◮ MAXE: The estimated link resources attached to a node

⇒ run time independent of substrate size and topology Removes symmetry from the problem to speed up the solution process

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 15 / 21

slide-46
SLIDE 46

OptCut

  • 8

9 11 13 15 17 19 21 23 25 27 29

Number of VNet nodes Solving time (s)

0.01 0.1 1 10 100 Basic formulation Formulation with order

  • Carlo F¨

urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 15 / 21

slide-47
SLIDE 47

Evaluation Parameters

Objective: Embed as many virtual resources as possible

Substrate

DC topologies (default FatTree with 432 hosts) Each physical element has 4 resource units

Requests

Randomized topologies (2-10 nodes, connection probability 0.15) Exponentially distributed duration with mean 10 Resource sum of all requests ≈ available substrate resources All Per-Clustering approaches are combined with SecondNet

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 16 / 21

slide-48
SLIDE 48

Experimental Pipeline

Embed. Algorithm Request sequence ADD REQ1 ADD REQ2 ADD REQ3 REM REQ1 ADD REQ4 STATE ... Substrate Topologies FatTree BCube DCell Pre- clustering Farhat OptCut Loco Unmodified Requests Modified Requests SNet SNet SNet LoCo SNet

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 17 / 21

slide-49
SLIDE 49

Performance Analysis

  • LoCo

OptCut* LoCo* Farhat* SecondNet 0.0 0.2 0.4 0.6 0.8 1.0

Node Utilization

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 18 / 21

slide-50
SLIDE 50

Performance Analysis

  • LoCo

OptCut* LoCo* Farhat* SecondNet 0.0 0.2 0.4 0.6 0.8 1.0

Node Utilization

All Pre-Clustering approaches improve the performance of Secondnet by factors > 1.5 But why is standalone LoCo in this scenario more preformant?

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 18 / 21

slide-51
SLIDE 51

Reason I: Good Scenario for LoCo

  • LoCo

OptCut* LoCo* Farhat* SNet 0.0 0.2 0.4 0.6 0.8 1.0

Size 10

Node Utilization

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 19 / 21

slide-52
SLIDE 52

Reason I: Good Scenario for LoCo

  • LoCo

OptCut* LoCo* Farhat* SNet 0.0 0.2 0.4 0.6 0.8 1.0

Size 11

Node Utilization

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 19 / 21

slide-53
SLIDE 53

Reason I: Good Scenario for LoCo

LoCo OptCut* LoCo* Farhat* SNet 0.0 0.2 0.4 0.6 0.8 1.0

Size 12

Node Utilization

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 19 / 21

slide-54
SLIDE 54

Reason I: Good Scenario for LoCo

  • LoCo

OptCut* LoCo* Farhat* SNet 0.0 0.2 0.4 0.6 0.8 1.0

Size 13

Node Utilization

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 19 / 21

slide-55
SLIDE 55

Reason I: Good Scenario for LoCo

  • LoCo

OptCut* LoCo* Farhat* SNet 0.0 0.2 0.4 0.6 0.8 1.0

Size 14

Node Utilization

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 19 / 21

slide-56
SLIDE 56

Reason I: Good Scenario for LoCo

  • LoCo

OptCut* LoCo* Farhat* SNet 0.0 0.2 0.4 0.6 0.8 1.0

Size 15

Node Utilization

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 19 / 21

slide-57
SLIDE 57

Reason II: Fragmented Residual Resources

  • 0.4

0.6 0.8 1 1.2 0.4 0.6 0.8 1.0

OptCut*

Load Node Utilization

  • 0.4

0.6 0.8 1 1.2 0.4 0.6 0.8 1.0

LoCo

Load

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 20 / 21

slide-58
SLIDE 58

What else is in the paper?

Description of the MetaTree Framework

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 21 / 21

slide-59
SLIDE 59

What else is in the paper?

Contains Represents Meta Info

  • Avail res.
  • Interference
  • Policy
  • ...

Has

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 21 / 21

slide-60
SLIDE 60

What else is in the paper?

Description of the MetaTree Framework Detailed description of LoCo

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 21 / 21

slide-61
SLIDE 61

What else is in the paper?

Require: VNet G = (V , E), M = {s} for some s ∈ V (G), P = (Γ(s)) while |P| > 0 do sort P (* decreasing link capacities *) choose u = P[0] (* next node to map *) map u (* forward checking *) map {u, v} ∀ v ∈ M, where {u, v} ∈ E(G) M = M ∪ {u} and P = P \ {u} end while if (embedding failed), backtrack on s

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 21 / 21

slide-62
SLIDE 62

What else is in the paper?

Description of the MetaTree Framework Detailed description of LoCo Concrete MIP formulations and evaluation

◮ Runtime comparison ◮ Impact of MAXE and MAXV Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 21 / 21

slide-63
SLIDE 63

What else is in the paper?

Constants:

Set of nodes: V (1) Set of edges: E ⊂ V × V (2) Weights: W : V ∪ E → R≥0 (3) Maximal node resources: MA XV (4) Maximal link resources: MA XE (5) Larger nodes: ρ : V → 2V (6)

Variables:

PC Node mapping: allocV : V × V → {0, 1} (7) Auxiliary variable: x : E × V → R≥0 (8) . . .

Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 21 / 21

slide-64
SLIDE 64

What else is in the paper?

Description of the MetaTree Framework Detailed description of LoCo Concrete MIP formulations and evaluation

◮ Runtime comparison ◮ Impact of MAXE and MAXV Carlo F¨ urst (TU Berlin) Virtual Network Embedding with Collocation November 12, 2013 21 / 21