1/30. Introduction. Key Concepts. GPSLA Scheduling. Early Experimentation. Conclusions and Future Work. A Green Scheduling Policy for Cloud Computing.
Introduction
Key Concepts
GPSLA Scheduling
Early Experimentation
Conclusions and Future Work
A Green Scheduling Policy for Cloud Computing J. Vilaplana1 , F. Solsona1 , J. Mateo1 , I. Teixid´o1 , J. Rius1 F.Abella2 1 Distributed
Computing Group, University of Lleida http://gcd.udl.cat/ 2 Santa Maria Hospital & IRB Lleida
ARMS-CC, July 2014
1/30 J. Vilaplana, F. Solsona, J. Mateo, I. Teixid´ o, J. Rius, F.Abella A Green Scheduling Policy for Cloud Computing
UdL
Introduction
Key Concepts
GPSLA Scheduling
Early Experimentation
Conclusions and Future Work
Table of Contents
1 Introduction 2 Key Concepts 3 GPSLA Scheduling 4 Early Experimentation 5 Conclusions and Future Work
2/30 J. Vilaplana, F. Solsona, J. Mateo, I. Teixid´ o, J. Rius, F.Abella A Green Scheduling Policy for Cloud Computing
UdL
Introduction
Key Concepts
GPSLA Scheduling
Early Experimentation
Conclusions and Future Work
Table of Contents
1 Introduction 2 Key Concepts 3 GPSLA Scheduling 4 Early Experimentation 5 Conclusions and Future Work
3/30 J. Vilaplana, F. Solsona, J. Mateo, I. Teixid´ o, J. Rius, F.Abella A Green Scheduling Policy for Cloud Computing
UdL
Introduction
Key Concepts
GPSLA Scheduling
Early Experimentation
Conclusions and Future Work
Introduction - Cloud Computing
Cloud computing is a model for enabling ubiquitous, convenient, on-demand network access to a shared pool of configurable computing resources
4/30 J. Vilaplana, F. Solsona, J. Mateo, I. Teixid´ o, J. Rius, F.Abella A Green Scheduling Policy for Cloud Computing
UdL
Introduction
Key Concepts
GPSLA Scheduling
Early Experimentation
Conclusions and Future Work
Introduction - Cloud Systems
Formed by multiple physical hosts Several VM are deployed in each host VM are created under demand
5/30 J. Vilaplana, F. Solsona, J. Mateo, I. Teixid´ o, J. Rius, F.Abella A Green Scheduling Policy for Cloud Computing
UdL
Introduction
Key Concepts
GPSLA Scheduling
Early Experimentation
Conclusions and Future Work
Introduction - Our Proposal
Green Preserving SLA (GPSLA) A power-aware scheduling policy algorithm to maintain SLA and minimize energy consumption in cloud systems.
6/30 J. Vilaplana, F. Solsona, J. Mateo, I. Teixid´ o, J. Rius, F.Abella A Green Scheduling Policy for Cloud Computing
UdL
Introduction
Key Concepts
GPSLA Scheduling
Early Experimentation
Conclusions and Future Work
Table of Contents
1 Introduction 2 Key Concepts 3 GPSLA Scheduling 4 Early Experimentation 5 Conclusions and Future Work
7/30 J. Vilaplana, F. Solsona, J. Mateo, I. Teixid´ o, J. Rius, F.Abella A Green Scheduling Policy for Cloud Computing
UdL
Introduction
Key Concepts
GPSLA Scheduling
Early Experimentation
Conclusions and Future Work
Key Concepts SLA: Service-Level Agreement Contract between a (Cloud) provider and a customer that specifies, in measurable terms, what services the provider will furnish. Power-aware Scheduling Minimize overall energy consumption by applying a defined scheduling policy. LP: Linear programming Solve an optimization problem defined by constraints, over a set of unknown real variables, along with an objective function (OF ) to be maximized or minimized, where constraints or the OF are linear. 8/30 J. Vilaplana, F. Solsona, J. Mateo, I. Teixid´ o, J. Rius, F.Abella A Green Scheduling Policy for Cloud Computing
UdL
Introduction
Key Concepts
GPSLA Scheduling
Early Experimentation
Conclusions and Future Work
Table of Contents
1 Introduction 2 Key Concepts 3 GPSLA Scheduling 4 Early Experimentation 5 Conclusions and Future Work
9/30 J. Vilaplana, F. Solsona, J. Mateo, I. Teixid´ o, J. Rius, F.Abella A Green Scheduling Policy for Cloud Computing
UdL
Introduction
Key Concepts
GPSLA Scheduling
Early Experimentation
Conclusions and Future Work
GreenC Scheduling
Provided a set of performance criteria, GPSLA aims to Minimize energy consumption Minimize job response time Return the assignment of tasks to VMs that minimize both power consumption and response time
10/30 J. Vilaplana, F. Solsona, J. Mateo, I. Teixid´ o, J. Rius, F.Abella A Green Scheduling Policy for Cloud Computing
UdL
Introduction
Key Concepts
GPSLA Scheduling
Early Experimentation
Conclusions and Future Work
GPSLA Scheduling
How to reduce energy consumption? GPSLA tries to assign as many tasks as possible to the most powerful VMs, so unused VM can then be turned off. GPSLA policy assumes A cloud system made up of V heterogeneous VMs A set of heterogeneous tasks T Each VMv has a specific amount of memory Mv
11/30 J. Vilaplana, F. Solsona, J. Mateo, I. Teixid´ o, J. Rius, F.Abella A Green Scheduling Policy for Cloud Computing
UdL
Introduction
Key Concepts
GPSLA Scheduling
Early Experimentation
Conclusions and Future Work
GPSLA - One node
The relative computing power (∆v ) is defined as the normalized score of a VM
So, given V VMs, ∆v = score of VMv
δ PV v
k=1 δk
, where
PV
k=1 ∆v
= 1. δv is the
The closer the relative computing power is to 1, the more likely it is that the requests will be mapped into such a VM
12/30 J. Vilaplana, F. Solsona, J. Mateo, I. Teixid´ o, J. Rius, F.Abella A Green Scheduling Policy for Cloud Computing
UdL
Introduction
Key Concepts
GPSLA Scheduling
Early Experimentation
Conclusions and Future Work
GPSLA - One node The objective function to be maximized is: max(
PV
v =1 tv ∆v )
Where the constraints are: s.t. :
PV
v =1
=T
tv ≤ Mv , ∀v ≤ V Where T is the total number of requests or tasks, and Mv the capacity or memory of a VM. 13/30 J. Vilaplana, F. Solsona, J. Mateo, I. Teixid´ o, J. Rius, F.Abella A Green Scheduling Policy for Cloud Computing
UdL
Introduction
Key Concepts
GPSLA Scheduling
Early Experimentation
Conclusions and Future Work
GPSLA - One node load-aware Underloaded When a VM is underloaded, its throughput will increase when more tasks are assigned to it. Overloaded When a VM reaches its maximum workload capacity, its throughput starts falling asymptotically towards 0.
This behaviour can be modelled using an Erlang distribution density function. 14/30 J. Vilaplana, F. Solsona, J. Mateo, I. Teixid´ o, J. Rius, F.Abella A Green Scheduling Policy for Cloud Computing
UdL
Introduction
Key Concepts
GPSLA Scheduling
Early Experimentation
Conclusions and Future Work
GPSLA - One node load-aware
Erlang is a continuous probability distribution with two parameters, α and λ, where α is the shape parameter and λ is the rate parameter. α−1
E (x; α, λ) = λe −λx (λx) (α−1)! ∀x, λ ≥ 0 These parameters depend on the VM characteristics. Erlang modelling parameters of each VM could be empirically obtained.
15/30 J. Vilaplana, F. Solsona, J. Mateo, I. Teixid´ o, J. Rius, F.Abella A Green Scheduling Policy for Cloud Computing
UdL
Introduction
Key Concepts
GPSLA Scheduling
Early Experimentation
Conclusions and Future Work
GPSLA - One node load-aware
16/30 J. Vilaplana, F. Solsona, J. Mateo, I. Teixid´ o, J. Rius, F.Abella A Green Scheduling Policy for Cloud Computing
UdL
Introduction
Key Concepts
GPSLA Scheduling
Early Experimentation
Conclusions and Future Work
GPSLA - One node load-aware The relative computing power ∆v of each VMv is weighted by its associated workload factor determined by an Erlang distribution. Now the objective function to be maximized is: P max( V v =1 tv ∆v E (tv ; α, λ)) Where the constraints are: P s.t. : V v =1 = T tv ≤ Mv , ∀v ≤ V Where T is the total number of requests or tasks, and Mv the capacity or memory of a VM. 17/30 J. Vilaplana, F. Solsona, J. Mateo, I. Teixid´ o, J. Rius, F.Abella A Green Scheduling Policy for Cloud Computing
UdL
Introduction
Key Concepts
GPSLA Scheduling
Early Experimentation
Conclusions and Future Work
GPSLA - One node load-aware heterogeneous tasks Each task ti has its Processing cost Pvi Pvi represents the execution time of task ti in VMv with respect to the execution time of ti in the less powerful VMv Mvi is defined as the amount of Memory allocated to task ti in VMv , where Mvi = Mv 0 i ∀v , v 0 ≤ V . The Boolean variable tvi represents the assignment of task ti to VMv , where tvi = 1 if t i is assigned to VMv , and tvi = 0 otherwise. 18/30 J. Vilaplana, F. Solsona, J. Mateo, I. Teixid´ o, J. Rius, F.Abella A Green Scheduling Policy for Cloud Computing
UdL
Introduction
Key Concepts
GPSLA Scheduling
Early Experimentation
Conclusions and Future Work
GPSLA - One node load-aware heterogeneous tasks The new linear model is: max(
PV
PT
s.t. :
PT
= Mvi ≤ Mv ∀v ≤ V
PV
v =1 (
i=1
v =1 tvi
i=1 Pvi tvi )∆v E (
PT
j=1 Pvj tvj ; α, λ))
= 1∀i ≤ T
Pvi tvi takes into account the computing cost of task ti in its assigned VMv . 19/30 J. Vilaplana, F. Solsona, J. Mateo, I. Teixid´ o, J. Rius, F.Abella A Green Scheduling Policy for Cloud Computing
UdL
Introduction
Key Concepts
GPSLA Scheduling
Early Experimentation
Conclusions and Future Work
Table of Contents
1 Introduction 2 Key Concepts 3 GPSLA Scheduling 4 Early Experimentation 5 Conclusions and Future Work
20/30 J. Vilaplana, F. Solsona, J. Mateo, I. Teixid´ o, J. Rius, F.Abella A Green Scheduling Policy for Cloud Computing
UdL
Introduction
Key Concepts
GPSLA Scheduling
Early Experimentation
Conclusions and Future Work
Early Experimentation Experimental results were obtained using the CPLEX mathematical optimizer, with 50 tasks (T = 50) and the following VM:
VM 1 2 3
∆v 50 40 10
Memory 3 2 2
Erlang Distribution α = 3, λ = 8 α = 2, λ = 8 α = 2, λ = 4
Table: VM Configurations
21/30 J. Vilaplana, F. Solsona, J. Mateo, I. Teixid´ o, J. Rius, F.Abella A Green Scheduling Policy for Cloud Computing
UdL
Introduction
Key Concepts
GPSLA Scheduling
Early Experimentation
Conclusions and Future Work
Early Experimentation - Without Erlang
All tasks get consolidated in VM 1, allowing to shutdown VM 2 and VM 3.
VM 1 50
VM 2 0
VM 3 0
However, VM 1 may have been overloaded.
22/30 J. Vilaplana, F. Solsona, J. Mateo, I. Teixid´ o, J. Rius, F.Abella A Green Scheduling Policy for Cloud Computing
UdL
Introduction
Key Concepts
GPSLA Scheduling
Early Experimentation
Conclusions and Future Work
Early Experimentation - With Erlang Next, the Erlang distribution for VM 1 is taken into account.
Erlang for VM 1 23/30 J. Vilaplana, F. Solsona, J. Mateo, I. Teixid´ o, J. Rius, F.Abella A Green Scheduling Policy for Cloud Computing
UdL
Introduction
Key Concepts
GPSLA Scheduling
Early Experimentation
Conclusions and Future Work
Early Experimentation - With Erlang Next, the Erlang distribution for VM 2 is taken into account.
Erlang for VM 2 24/30 J. Vilaplana, F. Solsona, J. Mateo, I. Teixid´ o, J. Rius, F.Abella A Green Scheduling Policy for Cloud Computing
UdL
Introduction
Key Concepts
GPSLA Scheduling
Early Experimentation
Conclusions and Future Work
Early Experimentation - With Erlang Next, the Erlang distribution for VM 3 is taken into account.
Erlang for VM 3 25/30 J. Vilaplana, F. Solsona, J. Mateo, I. Teixid´ o, J. Rius, F.Abella A Green Scheduling Policy for Cloud Computing
UdL
Introduction
Key Concepts
GPSLA Scheduling
Early Experimentation
Conclusions and Future Work
Early Experimentation - With Erlang VM placement distribution: VM 1 25
VM 2 17
VM 3 8
The scheduler now tries to put all the possible tasks in the most powerful VM, but when the performance starts decreasing significantly, it starts sending further tasks to the other available virtual machines. This behavior is more accurate as, although energy saving is a priority, we still want to preserve a certain degree of quality of service. 26/30 J. Vilaplana, F. Solsona, J. Mateo, I. Teixid´ o, J. Rius, F.Abella A Green Scheduling Policy for Cloud Computing
UdL
Introduction
Key Concepts
GPSLA Scheduling
Early Experimentation
Conclusions and Future Work
Table of Contents
1 Introduction 2 Key Concepts 3 GPSLA Scheduling 4 Early Experimentation 5 Conclusions and Future Work
27/30 J. Vilaplana, F. Solsona, J. Mateo, I. Teixid´ o, J. Rius, F.Abella A Green Scheduling Policy for Cloud Computing
UdL
Introduction
Key Concepts
GPSLA Scheduling
Early Experimentation
Conclusions and Future Work
Conclusions
Presented the GPSLA scheduling policy Minimize both energy consumption and response time Implemented using the CPLEX optimizer Early experimentation proves consistent in different scenarios Further experimentation is needed
28/30 J. Vilaplana, F. Solsona, J. Mateo, I. Teixid´ o, J. Rius, F.Abella A Green Scheduling Policy for Cloud Computing
UdL
Introduction
Key Concepts
GPSLA Scheduling
Early Experimentation
Conclusions and Future Work
Future Work
Expand the performance criteria (disk requirements, etc.) Perform further experimentation in different scenarios Implement GPSLA policy in a cloud simulator as CloudSim Scale the system and the workload Implement these strategies in OpenStack, a real cloud platform
29/30 J. Vilaplana, F. Solsona, J. Mateo, I. Teixid´ o, J. Rius, F.Abella A Green Scheduling Policy for Cloud Computing
UdL
Introduction
Key Concepts
GPSLA Scheduling
Early Experimentation
Conclusions and Future Work
Thank you
Thank you for your attention!
30/30 J. Vilaplana, F. Solsona, J. Mateo, I. Teixid´ o, J. Rius, F.Abella A Green Scheduling Policy for Cloud Computing
UdL