Efficient Autoscaling in the Cloud using Predictive Models for
EFFICIENT AUTOSCALING IN
THE CLOUD USING
PREDICTIVE MODELS FOR
WORKLOAD FORECASTING
Roy, N., A. Dubey, and A. Gokhale
4th IEEE International Conference on Cloud Computing (Cloud 2011)
1
Agenda
2
Introduction
Related Work
Challenge
Solution
Evaluation
Conclusion
Comment
Introduction (1/4)
3
Typically customers maintain SLAs with service providers for the QoS properties.
Failure to comply with satisfying these QoS metrics leads to a major loss of revenue in the form of decreased user base
Catering to the SLA while still keeping costs low is challenging for such enterprise systems
4
Introduction (2/4) – naïve method
Introduction (3/4)
5
A problem with such a resource allocation scheme
Is the chance of thrashing where due to frequent variation of workload, machines can be added and released on every sample
A desirable solution would require an ability to predict the incoming workload on the system and allocate resources a priori
Introduction (4/4)
6
autoscaling the resources in a cloud environment is not an easy and straightforward task.
(i) overheads related to state transition when number of resources are changed
(ii) ability to accurately predict future workload
(iii) compute the right number of resources required for the expected increase or decrease in workload.
Related Work (1/2)
7
(1) Heuristics-based virtual machine allocation and migration
Urgaonkar et. al. [2], 2008
VM, dynamic provisioning, Queueing model
Only a single VM can be run in a host
Wood et. al. [3], 2007
VM, dynamic migration define a unique metric based on the consumption data of the three resources to make the migration decision
CPU, network and memory
Cunha et. al. [4], 2007
Queueing model
Pricing Model that gives rewards for throughput to be within SLA limits and penalty for throughput going above
Related Work (2/2)
8
(2) Autonomic management of virtual computing environment using control-theoretic approaches:
Wang et. al. [6]
A load balancing controller
VMs are all load-balanced and the response time of the applications in all the VMs are the same
Moreno et. al. [7]
An architecture for elastic management of cluster-based services
Waheed et. al. [8]
Reactive algorithm to allocate resources to a cluster farm
Yang et. al [9]
Profiled based approach
Challenge (1/4)
9
Challenge 1: Workload Forecasting
Correctness of prediction
Releasing resources is easy, but..
Acquiring resources
Make a call on the cloud API which starts the acquisition process
The machine will be needed to boot up with the specified image
The application need to be started
Challenge (2/4)
10
Challenge 2: Identify Resource Requirement for
Incoming Load
The required number of resources is a function of
the number of customers
the nature of the application
the type of calls that each customer makes on the application
Challenge (3/4)
11
Challenge 3: Resource Allocation while Optimizing
Multiple Cost Factors
To optimize resource usage and/or minimize idle resources
define a time interval and change resources as many times as possible as workload changes.
In the limit, this interval could be made infinitesimally small and resources are changed continuously in accordance with the change in load
12
Challenge (4/4)
Obviously, such as scheme is not possible
the overhead in allocating a resource
scaling up or down resources also involves cost and needs to be optimized
Solution (1/9)
13
Control theory offers a promising methodology to address the challenges
Solution (2/9)
14
1. For every future time step, it computes the cost of selecting each possible resource allocation
2. To compute the cost of a particular allocation, it uses Algorithm 1 to compute the estimated response time for that particular machine configuration
3. Once the response time is calculated, it is used to calculate the cost of the allocation which is a combination of
how far the estimated response time is from the SLA bounds (SLA violation) cost of leasing additional machines and also a cost of re-configuration
Solution (3/9)
15
A. Workload Prediction
Authors used a second order autoregressive moving average method (ARMA) filter for the workload
( t
1 )
( t )
( t
1 )
( 1
(
))
( t
2 )
The value for the variables β and γ are given by the values 0.8 and 0.15
ARMA
16
Autoregressive Model (AR)
a model depends on the level of the lagged observations
For example, if we observe a high realisation of GDP we would expect that the GDP in the next few periods are high as well x t
c
i p
1
i x t
i
t
ARMA
17
Moving Average Model (MA)
model that the observations of a random variable at time t are not only affected by the shock at time t, but also the shocks of prior periods
Ex. if we observe a negative shock to the economy, say,
9/11, then we would expect that the negative effect affects the economy also for the near future.
t
q x
i
1
i
t
i
t
ARMA
18
Autoregressive Moving Average Model
combine both models we get a ARMA(p,q) model x t
c
t
i p
1
i x t
i
i q
1
i
t
i
ARMA models are widely used for prediction of economic and industrial time series
19
Solution (4/9)
Solution (5/9)
20
B. Performance Model
The next challenge we resolve is identifying resource requirements for the predicted workload
The workload used in this work is the number of users currently in the system.
It also depends upon what each user does.
In prior work [20] we have used Customer Behavior
Modeling Graphs (CBMG) (?) to model the overall behavior of customers
Solution (6/9)
21
A CBMG is built from a log of previous customer behavior and computes the probability of a typical user to visit each page
Using this information, we can calculate the number of visits to a single page from the total number of customers in the system.
The number of visits to each page helps in calculating the average load on each page.
22
Solution (7/9)
Solution (8/9)
23
C. Optimizing Resource Provisioning
The intuition is to identify the right number of time intervals
Our solution works on look-ahead optimization
iteratively solves an optimization problem, Cost opt , starting from t
0
Solution (9/9)
24
The next challenge is the choice of the look-ahead period.
A small look-ahead period will neglect trends
A very large period will increase computational complexity
The actual algorithm is not described here
because the implementation requires recursive data structures
is difficult to describe in the limited space available.
Evaluation (1/7)
25
Cost Function
26
Evaluation (2/7)
Just in time Resource Allocation
the weights on each component of the cost function is the same
27
Evaluation (3/7) -- Resource Usage under Different Cost Priorities
1) SLA violation against Resource Cost
The ratio of SLA penalty to machine cost is varied from
4 : 1 to 1 : 13
28
Evaluation (4/7) -- Resource Usage under Different Cost Priorities
29
Evaluation (5/7) -- Resource Usage under Different Cost Priorities
30
Evaluation (6/7) -- Resource Usage under Different Cost Priorities
2) Including the Cost of Reconfiguration
31
Evaluation (7/7) -- Resource Usage under Different Cost Priorities
Conclusion
32
this paper describes a look-ahead resource allocation algorithm based on model predictive control
predicts future workload
adjusts resources allocated to users ahead-of-time
Comments
33
The detail of the model in the paper is too simple
I cannot understand why the authors did these evaluations
The paper use control theory and it seems to have a good prediction of workload
Something in 3 challenges
the overhead of allocating resources
the prediction interval
the costs of SLA violation, reconfiguring machine….
