Upload
salam-ismaeel
View
333
Download
0
Embed Size (px)
Citation preview
Multivariate Time Series ELM for Cloud Data
Centre Workload Prediction Consumption
Salam Ismaeel and Ali Miri
Department of Computer Science
Ryerson University 350 Victoria Street, Toronto, Ontario, Canada M5B 2K3
and
July 21th, 2016
Agenda
• Problem Definition
• Work Contribution
• Workload Prediction System
• Experimental Results
• Conclusions
2
Problem Definition
3
CDC delivering computing resources on-
demand over the Internet
Data centres power (2% of global
electricity usage), expected to double by
2020.
Data centre electricity use, [2015 Alliance Trust]
A Google’s data centre, [Google,2012]
4
Power Management Techniques
- Dynamic
- Software Level
-Cloud Data Centre
Problem Definition
Dynamic Consolidation of VM, [Beloglazov,2014]
Dynamic consolidation of VMs
Live migration and/or optimal
placement of new VMs.
5
Find a way to make predictions that take into account both user and VM variations.
Overcome the problem of multivariate time varying VM requests
Eliminate the restrictions about observation window size and number of VM clusters
Able to detect when prediction is likely to be incorrect and how we can overcome the
problem?
How to predict the CDC workload? How we can predict future VM requests?
Problem Definition
Agenda
• Problem Definition
• Work Contribution
• Workload Prediction System
• Experimental Results
• Conclusions
6
7
We use clustering not only on VM requests, but also on user requests. This can
result in proper filtering of unexpected VM requests caused by unpredictable
users' actions.
We overcome the problem of time varying VM requests, depending on the actual
service demand.
Our proposed multivariate time series ELM represents an online sequential
framework which is able to eliminate the restrictions about observation window
size and number of VM clusters (inputs) for the ELM predictor.
Work Contribution
Agenda
• Problem Definition
• Work Contribution
• Workload Prediction System
• Experimental Results
• Conclusions
8
9
Workload Prediction System
Proposed Prediction Framework
Workload prediction (future VM request prediction) based on available historical data; VM and
User cluster algorithms, Current state of the CDC; effective prediction window size:
Clustering Process
Prediction Process
10
Off-line clustering
To create a set of clusters for different
types of VMs and users form long term
historical data.
Trace decomposer
Responsible for mapping each request
received during an observation window.
VM and User
Behaviour
Prediction
Window size
Historical
Workload
Off line
clustering
Traces
Decomposer
Clustering Process
No. of
clusters
Cluster
centers
Workload
Prediction
Prediction Process
VM Request
Gathering
VM and User
Behaviour
Prediction
Window size
Historical
Workload
Off line
clustering
Traces
Decomposer
Clustering Process
No. of
clusters
Cluster
centers
Workload
Prediction
Prediction Process
VM Request
Gathering
Proposed Prediction Framework, [Salam, 2016]
Workload Prediction System
11
Off-line clustering
Trace decomposer
VM and User
Behaviour
Prediction
Window size
Historical
Workload
Off line
clustering
Traces
Decomposer
Clustering Process
No. of
clusters
Cluster
centers
Workload
Prediction
Prediction Process
VM Request
Gathering
Workload Prediction System
User and VM Behaviours
Comprehensive workload models must consider both VMs users behaviour to reflect realistic
conditions by excluding unwanted VMs or users form workload estimation process.
12
Off-line clustering
Trace decomposer
User and VM Behaviours
VM and User
Behaviour
Prediction
Window size
Historical
Workload
Off line
clustering
Traces
Decomposer
Clustering Process
No. of
clusters
Cluster
centers
Workload
Prediction
Prediction Process
VM Request
Gathering
Workload Prediction System
Historical Workload
- updated periodically
- predict the next period VM request for each observation
- used to calculate centres of clusters form time to time using long term observations
13
Off-line clustering
Trace decomposer
User and VM Behaviours
Historical Workload
VM and User
Behaviour
Prediction
Window size
Historical
Workload
Off line
clustering
Traces
Decomposer
Clustering Process
No. of
clusters
Cluster
centers
Workload
Prediction
Prediction Process
VM Request
Gathering
Workload Prediction System
VM Request Gathering
includes types of monitoring which can help in detecting and tracing the variations or failure of
resources and applications during an observation window
14
Off-line clustering
Trace decomposer
User and VM Behaviours
Historical Workload
VM Request Gathering
VM and User
Behaviour
Prediction
Window size
Historical
Workload
Off line
clustering
Traces
Decomposer
Clustering Process
No. of
clusters
Cluster
centers
Workload
Prediction
Prediction Process
VM Request
Gathering
Workload Prediction System
Prediction Window size
Is the time period for which the workload needs to be predicted to decide whether PMs need to be
switched to sleep mode.
It is totally depends on the configuration of CDC, specially the server hardware, and its values
effect on workload prediction section.
Workload Prediction System
15
Off-line clustering
Trace decomposer
User and VM Behaviours
Historical Workload
VM Request Gathering
Prediction Window size
VM and User
Behaviour
Prediction
Window size
Historical
Workload
Off line
clustering
Traces
Decomposer
Clustering Process
No. of
clusters
Cluster
centers
Workload
Prediction
Prediction Process
VM Request
Gathering
Workload PredictionMultivariate time series ELM predication algorithm has been improved for VM request
forecasting. Deal with
- time varying VM requests,
- eliminates restrictions on observation window sizes and number of VM clusters
Agenda
• Problem Definition
• Work Contribution
• Workload Prediction System
• Experimental Results
• Conclusions
16
10 hours of Google data, total number of request was 1,029,342, each of which has an
associated value of CPU and Memory.
Experimental Results
17
Fig. 2. SSD vs Number of cluster for (a) VMs (b) Users
We used the same size of observation windows and prediction windows, 800 and 60 second
respectively. we chose to use number of hidden neuron = 100 and 1/(regulation parameter) =
1/5000 for the proposed predictor
Experimental Results
17Fig.4. RMSE comparisons of different predictive approaches
Agenda
• Problem Definition
• Work Contribution
• Workload Prediction System
• Experimental Results
• Conclusions
20
Conclusion
• We have proposed a real-time CDC workload prediction framework that can be usedfor a better energy conservation strategy.
• A key component of this framework is a new multivariate time series ELMpredication algorithm for VM request forecasting.
• The improve ELM can deal with the problem of time varying VM requests, andeliminates any restrictions on observation window sizes and the number of VMclusters for the ELM predictor.
• The framework is based on efficient use of historical VM request, user clusteralgorithms, the current state of the data centre and an effective prediction windowsize.
21