Dynamic Resource Allocation in Clouds: Smart Placement with Live Migration Makhlouf Hadji Ingénieur de Recherche
[email protected]
Avec : Djamal Zeghlache (TSP)
[email protected]
FONDATION DE COOPERATION SCIENTIFIQUE
2
NOS VALEURS
❶
Excellence ❷ Souplesse
- Talents - Résultats - Echanges
- Fonctionnement - Partenariat - Projets ❸
Rigueur
- Propriété Intellectuelle - Confidentialité - Exécution 3
PROGRAMMES DE R&D Ingénierie numérique des Systèmes et Composants
4
PROJETS R&D Projets opérationnels depuis le lancement
9 15,5
Projets Opérationnels M€ Financement Industriel
29 ETP/an sur 3 ans 110 42 Partenaires Industriels 12 Partenaires Académiques
Thèses
5
I- Smart Placement in Clouds
Smart Placement in Clouds VM Placement problem Problem: Based on allocating and hosting N VMs on a physical infrastructure of X Serveurs, what is the best manner to optimally place workloads to minimize different infrastructure costs ?
Non optimal placement ESX 1
ESX 2…
ESX N
VMs demand management
VMs placement strategy
Cloud End-Users
???
Optimal placement ESX 1
ESX 2…
ESX N ESX 1
Benefits
Resources optimization,
Minimization of infrastructure costs,
Energy consumption optimization.
Challenges of the problem: Exponential number of cases to enumerate.
7
ESX 2…
ESX N
Infrastructure servers
Define the best strategy to place VMs workloads leading to optimally reduce infrastructure costs.
Smart Placement in Clouds
French Providers Point of View
8
Smart Placement in Clouds
Due to fluctuations in users’ demands, we use Auto-Regressive (AR(k)) process, to handle with future demands:
Gestion de la demande de VM
Utilisateurs des services Cloud
Forcasting & Scheduling
k
d t i d t i t
large small
i 1
ESX 1
Problem Complexity : NP-Hard Problem: One can construct easily a plynomial reduction from the NP-Hard notary problem of the BinPacking.
ESX 2…
ESX N
Infrastructure serveurs
9
Smart Placement in Clouds Mathematical formulation: N
Formulation as ILP: The corresponding mathematical model is an Integer Linear Programming: difficulties to characterize the convex hull of the considered problem and the optimal solution.
I
N
I
min Z ij yij Pj xij i 1 j 1
i 1 j 1
Subject To : xij Cij yij , j I , i 1, N N
x i 1
ij
d j , j I
xij N , i, j 1 if VM j is hosted in server i yij 0 else.
10
Smart Placement in Clouds Minimum Cost Maximum Flow Algorithm
Instance i
(2; 0,23)
S
T
Legend: (capacity; cost)
11
Smart Placement in Clouds Small Instance
Minimum Cost Maximum Flow Algorithm
(2; 0,23)
Medium Instance
S
T (2; 0,23)
12
Smart Placement in Clouds Simulation Tests: Case of (0;1) Random Costs
Random Hosting Costs Scenario We consider (0; 1) Random hosting costs between each couple of vertices (a, b), where a is a fictif node, and b is a physical machine (server).
13
Smart Placement in Clouds Simulations Tests: Case of Inverse Hosting Costs:
Inverse Hosting Costs Scenario We consider inversed hosting costs function between each couple of vertices (a, b), where a is a fictif node, and b is a physical machine:
Where
1 g ab if Cab 0, otherwise g ab f (Cab ) Cab represents the available capacity on the considered arc. f est une fonction non nulle.
14
II- Live Migration of VMs
Live Migration of VMs
Migration process:
Xen
ESX
KVM
Hyper-V
16
Live Migration of VMs Polyhedral Approachs
Polytops, faces and facets M in c x j
j
Subject To constraints:
1x 1
aij x jbi , i1,...,m xi 0,1 , i 1,..., n
x* facets face
2x 2 20
Live Migration of VMs Polyhedral Approachs
Some Valid Inequalities of our Problem:
Decision Variables:
Z ijk 1 if A VM k is migrated from i to j (0 else). Prevent backword migration of a VM:
Z ijk Z jlk ' 1 Server’s destination uniqueness of a VM migration: m'
Z
j 1, j i
1
Servers’ power consumption limitation constraints: m'
qi
ijk
Etc…
i 1 k 1
pk zijk p j ,max p j ,current 1 y j
18
Live Migration of VMs max M
m'
P i 1
i ,id l
qi
m'
m'
i 1
j 1 k 1
yi p 'k zijk
Polyhedral Approachs
Subject To : zijk z jlk ' 1, i 1, m, j i, m', k 1, qi , k ' 1, q j , l 1, m', l j , k k ' m'
z
j 1, j i m'
1
ijk
qi
p i 1 k 1 m'
k
zijk Pj , max Pj ,cu rren t 1 y j
qi
z j 1 k 1
ijk
qi yi
m' P j ,cu rren t m' j 1 yi m' Pj , max i 1 zijk t k T0 1 if VM k is migrated from i to j, zijk 0 otherwise 1 if server i is idle yi 0 otherwise
19
Live Migration of VMs Polyhedral Approachs Number of used servers when taking into account Migrations
20
Convergence Time (in seconds) of Migration Algorithm:
21
Live Migration of VMs Polyhedral Approachs Percentage of Gained Energy when Migration is Used
5
10
15
20
25
30
10
35,55
36,59
00
00
00
00
20
27,29
34,00
35,23
38,50
00
00
30
17,48
27,39
35,21
40,32
41,89
36,65
40
16,77
18,85
22,02
32,31
39,90
40,50
50
10,86
16,17
19,85
22,30
39,20
36,52
60
08,63
14,29
18,01
22,13
25,15
30,68
70
08,10
14,00
14,86
15,90
22,91
23,20
80
07,01
10,20
10,91
15,34
17,02
21,60
90
06,80
09,32
10,31
14,70
16,97
19,20
100
05,90
07,50
08,40
12,90
16,00
14,97
22
Thank you
