Continuous Resources Allocation in Internet Data Centers

1
Continuous Resources Allocation in Internet Data
Centers

• Youssef Hamadi
• Microsoft Research Cambridge

2
Internet Data Center
Website hosting

• Total availability in hosting
• 24/24, 7/7
• Power plant
• Secured access

3
Outline

• Constraint Programming overview
• Problem modelling
• Online problem solving
• Experiments
• Advanced Reservation in Grid Infrastructures
• Conclusion

4
Constraint Programming

• Problem Variables Constraints
• Variables Xlb..ub
• Constraints
• Input/output variables
• Events domain-reduction
• Action
• Logic, e.g., XYZ
• Operational, space reduction
• Algorithmic, complexity
• Resolution Constraint Propagation Speculative
  search

5
Constraint Programming
THINK
GUESS

Constraint propagation
Fix point
Speculative search
6
Constraint Programming
THINK
GUESS

7
Constraint Programming
8
Internet Data Center
Website hosting

• Total availability in hosting
• 24/24, 7/7
• Power plant
• Secured access

9
Problem Modelling
Internet Data Center
SE
Switches
SR
SR
SR
SR
SR
SR
SR
SR
SR
SR
SR
SR
Servers
C1
C3
C2
C4
C6
C5
C7
C9
C8
C10
C12
C11
C1
C3
C2
C4
C6
C5
C7
C9
C8
C10
C12
C11
10
Problem Modelling
Internet Data Center set of limitations
Mesh switch BSM_limit
Edge switch BSE_limit
Rack switch BSR_limit
Server CPU_limit Speed_limit Memory_limit
Storage_limit Disk_bandwidth_limit BC_limit
11
Problem Modelling
Multi-tier application
Internet
Web servers
Application servers
Databases
12
Problem Modelling
Multi-tier application set of requirements
Internet
Web servers
Process CPU_charge Speed_charge Memory_charge
Storage_charge Disk_bandwidth_charge
Bandwidth_charge
Application servers
Databases
13
Optimal Resource Allocation
Internet Data Center
Limitations
Requirements
Multi-tier application
14
Modelling

• Variables?
• Values?
• Constraints?

15
Modelling, variables

• Switch capacities
• Sm.in/Sm.out
• Sri.in/Sri.out
• Sei.in/Sei.out

N

• Network Capacities
• Si.in/Si.out N
• Allocated process
• Process N
• Tier1, Tier2, Tier3 boolean

16
Modelling, variables

• G (X, E)
• X, set of constrained processes
• E, comm. topology
• Allocated server
• Server N

17
Modelling, constraints

• 1. Static capacity filtering
• IDCs servers keep compatible processes
• ?Sk, ?Pk, k?Sk.Process, iff,
• Sk.CPU Pk.CPU,
• Sk.Speed Pk.Speed,
• Sk.Memory Pk.Memory,
• Sk.Storage Pk.Storage,
• Sk.DiskSpeed Pk.DiskSpeed,

18
Modelling, constraints

• 1. Static capacity filtering
• Application processes keep compatible servers
• ?Pk, ?Sk, k?Pk.Server, iff,
• Pk.CPU
• Pk.Speed
• Pk.Memory
• Pk.Storage
• Pk.DiskSpeed

19
Modelling, constraints

• 2. Symmetrical referencing

ServerSi
ProcessP1 Tier11 Tier20 Tier30
20
Modelling, constraints

• 3. Mutual exclusion
• ? Sk, all_different(Sk.Process)

21
Modelling, constraints

• 4. Tier propagation
• From a hosted process to the associated tier
  variables

P0, P1, P2, P3, P4, P5, P6, P7 tier1 1
1 0 0 0 0 0 0 tier2
0 0 1 1 1 0 0 0
tier3 0 0 0 0 0 1 1
1
Sk.Tier1 tier1Sk.Process Sk.Tier2
tier2Sk.Process Sk.Tier3 tier3Sk.Process
ProcessP0, P1 Tier11 Tier20 Tier30
22
Modelling, constraints

• 5. Bandwidth capacities, For each Rack Srk,
• Srk.in
• Srk.out

client
bc01
n1 2
b
a
bc12
n2 3
c
d
e
bc23
a
c
d
f
g
h
n3 3
23
Modelling, soft constraints

• Cost function

b01 b12 b23

2
4
6
24
Extended Modelling

• Symmetry break in a Rack switch

Usually, Sj ? Sk If Sj.Process ?
Sk.Process, Sj.Process according to monotonic properties of filtered
domains)
a,b,c
a,b,c
a,b,c
a b c
25
Extended Modelling

• Symmetry break in the application

? Pi, Pj at the same tier, ?(Pi) ?(Pj) Ordering
constraint Pi.Server
26
Optimal Resource Allocation
Internet Data Center
Limitations
Constraint Solver
Requirements (SLAs)
27
Outline

• Constraint Programming an overview
• Problem modelling
• Online problem solving
• Experiments
• Conclusion

28
The problem is Online (I)

• Web site annual load (Arlitt al. ACM-TOIT01)

request
time
Advertising campaign
Christmas
Competitor break
bank holiday
29
The problem is Online (II)

• Evolution of website traffic

Web servers
Static content
Application servers
Dynamic content
Databases
30
The problem is Online (III)

• Component failures

31
Online resource allocations in IDC
The major part of the users use personalization!
32
Online resource allocations in IDC
33
Constraint Programming
P
Problems
m
CP modelling
DechterDechter88
34
Constraint Programming
Component failure
Se2.capacity_in 0 Se2.capacity_out 0
Applications lifecycle reduction
Pi.Server -1
35
Online Architecture
IDC
Learning
Phase transition parameters
Search statistics
Search parameters
Add/remove constraints
Search control
Load /topology variations
monitoring
results
Search module
Heuristics
Setup
Contract negotiation, SLAs,
Cost
solve
management
Feasibility, Solve, State,
36
Experiments

• IDC with 1024 servers,
• 8 Edge switches, 8 Rack switches, 16 servers/Rack
• 3-tiers application
• (3,1,1), (3,2,2)
• /- (50000 constraints, 5000 variables)

37
Advanced Reservation in Grid infrastructures

• Gridline Project
• Microsoft Research Cambridge
• York University

38
Advanced Reservations

• Definition The process of requesting various
  resources for use at a later time.
• GGF Definition "An advance reservation is a
  possibly limited or restricted delegation of a
  particular resource capability over a defined
  time interval, obtained by the requester from the
  resource owner through a negotiation process."
• Example of resource capabilities number of
  processors, amount of memory, disk space,
  software licences, network bandwidth, etc.

39
Advanced Reservations

• Gridline
• Goal Maximize the utility of some resource
  broker
• How Compute an optimal subset of customers

40
Advanced Reservations

• Order
• Start/End date,
• Proposed Price,
• Proposed Penalty,
• QoS

41
Sample Instance of TKP
Uniform Capacity 10
Bid1 6 units, 11
Bid2 6 units, 10
Bid3 5 units, 20
t2
t3
t4
t5
t1
42
Time
43
Relative Quality
44
Conclusion
45
Conclusion

• Online Architecture for resources managements in
  IDCs.
• Constraint Programming modelling periodically
  refined.
• Future work
• Extend cost function
• real price of migrations, cooling requirements

46
Conclusion

• Gridline (joint work with York University)
• Resource allocation for the Grid
• Advanced reservation allocation CP-AI-OR05
• Workflow scheduling.
• http//www.cs.york.ac.uk/aig/constraints/Grid/

47
Conclusion

• Gridline (joint work with York University)
• Resource allocation for the Grid
• Advanced reservation allocation CP-AI-OR05
• Workflow scheduling.
• http//www.cs.york.ac.uk/aig/constraints/Grid/